diff --git a/examples/kumo-example-notebooks/structured-data-models/README.md b/examples/kumo-example-notebooks/structured-data-models/README.md
index 9de5abf4b..c63b0eca0 100644
--- a/examples/kumo-example-notebooks/structured-data-models/README.md
+++ b/examples/kumo-example-notebooks/structured-data-models/README.md
@@ -14,6 +14,13 @@ Client examples require an existing endpoint; they do not provision one.
### Cross-industry
+#### Notebooks
+
+| Example | Deployment | Scope |
+| --- | --- | --- |
+| [Improve KumoRelational performance on imbalanced data, with a fraud example](relational/cross-industry/notebooks/local_structured_data_models_imbalanced_classification_tuning.ipynb) | Local | Model workflow |
+| [Improve KumoRelational regression accuracy, with a spend-forecasting example](relational/cross-industry/notebooks/local_structured_data_models_regression_tuning.ipynb) | Local | Model workflow |
+
#### SDK code examples
| Example | Deployment | Scope |
diff --git a/examples/kumo-example-notebooks/structured-data-models/relational/cross-industry/notebooks/local_structured_data_models_imbalanced_classification_tuning.ipynb b/examples/kumo-example-notebooks/structured-data-models/relational/cross-industry/notebooks/local_structured_data_models_imbalanced_classification_tuning.ipynb
new file mode 100644
index 000000000..6bd4280c9
--- /dev/null
+++ b/examples/kumo-example-notebooks/structured-data-models/relational/cross-industry/notebooks/local_structured_data_models_imbalanced_classification_tuning.ipynb
@@ -0,0 +1,1676 @@
+{
+ "cells": [
+ {
+ "cell_type": "markdown",
+ "id": "24b42002",
+ "metadata": {},
+ "source": [
+ "# Improve KumoRelational performance on imbalanced data\n",
+ "\n",
+ "KumoRelational, a relational foundation model (RFM), predicts from labeled in-context examples. This notebook demonstrates four strategies to improve performance and measures the lift of each, using an imbalanced classes fraud detection use case as the example: which accounts will be confirmed as fraud within 30 days?\n",
+ "\n",
+ "| Strategy | What changes |\n",
+ "| --- | --- |\n",
+ "| 1. Rebalance the context | Include all positives; test class ratios, context size and positive repetition. |\n",
+ "| 2. Ensemble contexts | Keep the positives fixed, vary the negatives, average the predictions. |\n",
+ "| 3. Tune sampling | Recent entities, most recent neighbors, neighbor count. |\n",
+ "| 4. Tune the threshold | Pick a cutoff on validation data for each configuration. |\n",
+ "\n",
+ "Runs locally with Python 3.12+ and [direct KumoRelational](https://nvidia.github.io/structured-data-models/api/generated/sdm.models.KumoRelational.html), in about 20 minutes on CPU. Data is synthetic; metrics are **demonstration only**."
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "b89d6820",
+ "metadata": {},
+ "source": [
+ "## Install and configure\n",
+ "\n",
+ "Sampling runs on CPU even when the model uses a GPU, so it always needs a [pyg-lib wheel](https://pytorch-geometric.readthedocs.io/en/latest/install/installation.html) matching your PyTorch version; the schema diagram needs [Graphviz](https://graphviz.org/download/). See [installation](https://nvidia.github.io/structured-data-models/install.html)."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 1,
+ "id": "ceb339d5",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:48:29.509845Z",
+ "iopub.status.busy": "2026-11-19T01:48:29.509726Z",
+ "iopub.status.idle": "2026-11-19T01:48:30.347347Z",
+ "shell.execute_reply": "2026-11-19T01:48:30.346745Z"
+ }
+ },
+ "outputs": [],
+ "source": [
+ "%pip install structured-data-models pandas pyarrow scikit-learn matplotlib graphviz"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "d0325b98",
+ "metadata": {},
+ "source": [
+ "## Set the parameters\n",
+ "\n",
+ "Accounts are split by observation date into a context pool, a validation window for choosing settings and thresholds, and a test window for reporting."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 2,
+ "id": "36dfda2d",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:48:30.348994Z",
+ "iopub.status.busy": "2026-11-19T01:48:30.348870Z",
+ "iopub.status.idle": "2026-11-19T01:48:30.351487Z",
+ "shell.execute_reply": "2026-11-19T01:48:30.350968Z"
+ }
+ },
+ "outputs": [],
+ "source": [
+ "SEED = 7\n",
+ "N_ACCOUNTS = 100_000\n",
+ "FRAUD_RATE = 0.007\n",
+ "START = \"2026-01-01\"\n",
+ "HISTORY_DAYS = 90\n",
+ "HORIZON_DAYS = 30\n",
+ "VALIDATION_START_DAY = 250\n",
+ "TEST_START_DAY = 310\n",
+ "\n",
+ "BASELINE_CONTEXT_SIZE = 5_000\n",
+ "LARGE_CONTEXT_SIZE = 10_000\n",
+ "FRAUD_REPEAT = 10\n",
+ "NUM_ESTIMATORS = 5\n",
+ "RECENT_CONTEXT_DAYS = 90\n",
+ "FIRST_HOP_NEIGHBORS = 8\n",
+ "SECOND_HOP_NEIGHBORS = 4\n",
+ "NEIGHBOR_OPTIONS = [16]\n",
+ "\n",
+ "DEFAULT_THRESHOLD = 0.5\n",
+ "TARGET_PRECISION = 0.5\n",
+ "OUTPUT_PARENT = None\n",
+ "CLEANUP = False"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "55344016",
+ "metadata": {},
+ "source": [
+ "Import the libraries and create an output directory for this run."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 3,
+ "id": "8a4739ef",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:48:30.352753Z",
+ "iopub.status.busy": "2026-11-19T01:48:30.352661Z",
+ "iopub.status.idle": "2026-11-19T01:48:32.970254Z",
+ "shell.execute_reply": "2026-11-19T01:48:32.969851Z"
+ }
+ },
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "Device: cpu | Results: /var/folders/d1/l_l86lcn0jb9jls_r4k74tbc0000gn/T/fraud-context-tuning-c36lf3ed\n"
+ ]
+ }
+ ],
+ "source": [
+ "import tempfile\n",
+ "from pathlib import Path\n",
+ "\n",
+ "import matplotlib.pyplot as plt\n",
+ "import numpy as np\n",
+ "import pandas as pd\n",
+ "import sdm\n",
+ "import torch\n",
+ "from IPython.display import display\n",
+ "from sklearn.metrics import average_precision_score, f1_score, precision_recall_curve, precision_score, recall_score, roc_auc_score\n",
+ "\n",
+ "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
+ "rng = np.random.default_rng(SEED)\n",
+ "workspace = tempfile.TemporaryDirectory(prefix=\"fraud-context-tuning-\", dir=OUTPUT_PARENT, delete=False)\n",
+ "run_dir = Path(workspace.name)\n",
+ "print(\"Device:\", device, \"| Results:\", run_dir)"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "f1986d35",
+ "metadata": {},
+ "source": [
+ "## Generate synthetic example data\n",
+ "\n",
+ "Run the next cell to create five tables: customers, accounts, transactions, merchants and devices. You can skim it; the workflow starts at **Build the relational data**.\n",
+ "\n",
+ "Fraud accounts tend to show a burst of activity just before their observation date, often at high-risk merchants or on shared or emulated devices, and belong to newer customers. Some legitimate accounts look similar, and the fraud pattern drifts in early May. Accounts carry an open date and customers a join date, so the sampler can exclude records created after each observation date."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 4,
+ "id": "678a7276",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:48:32.971413Z",
+ "iopub.status.busy": "2026-11-19T01:48:32.971314Z",
+ "iopub.status.idle": "2026-11-19T01:48:33.348974Z",
+ "shell.execute_reply": "2026-11-19T01:48:33.348660Z"
+ }
+ },
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "
\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " rows | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | customers | \n",
+ " 75000 | \n",
+ "
\n",
+ " \n",
+ " | accounts | \n",
+ " 100000 | \n",
+ "
\n",
+ " \n",
+ " | transactions | \n",
+ " 2821422 | \n",
+ "
\n",
+ " \n",
+ " | merchants | \n",
+ " 2000 | \n",
+ "
\n",
+ " \n",
+ " | devices | \n",
+ " 76500 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
"
+ ],
+ "text/plain": [
+ " rows\n",
+ "customers 75000\n",
+ "accounts 100000\n",
+ "transactions 2821422\n",
+ "merchants 2000\n",
+ "devices 76500"
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ },
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "Fraud rate: 0.69% (about 1 in 145)\n"
+ ]
+ }
+ ],
+ "source": [
+ "N_CUSTOMERS = int(N_ACCOUNTS * 0.75)\n",
+ "N_MERCHANTS = 2_000\n",
+ "N_SHARED_DEVICES = 1_500\n",
+ "DRIFT_DAY = 120\n",
+ "\n",
+ "merchants = pd.DataFrame({\n",
+ " \"merchant_id\": np.arange(N_MERCHANTS),\n",
+ " \"risk_score\": rng.beta(1.5, 12, N_MERCHANTS).round(3),\n",
+ " \"is_online\": (rng.random(N_MERCHANTS) < 0.4).astype(float),\n",
+ " \"avg_ticket\": rng.lognormal(3.6, 0.6, N_MERCHANTS).round(2),\n",
+ "})\n",
+ "risky_merchants = {\n",
+ " \"early\": merchants.merchant_id[(merchants.risk_score > merchants.risk_score.quantile(0.9)) & (merchants.is_online == 0)].to_numpy(),\n",
+ " \"late\": merchants.merchant_id[(merchants.risk_score > merchants.risk_score.quantile(0.9)) & (merchants.is_online == 1)].to_numpy(),\n",
+ "}\n",
+ "everyday_merchants = merchants.merchant_id[merchants.risk_score <= merchants.risk_score.quantile(0.9)].to_numpy()\n",
+ "\n",
+ "customers = pd.DataFrame({\n",
+ " \"customer_id\": np.arange(N_CUSTOMERS),\n",
+ " \"customer_age_years\": rng.integers(18, 85, N_CUSTOMERS).astype(float),\n",
+ " \"joined_at\": pd.Timestamp(START) - pd.to_timedelta(rng.gamma(3.0, 400, N_CUSTOMERS).round(), unit=\"D\"),\n",
+ " \"kyc_tier\": rng.integers(1, 4, N_CUSTOMERS).astype(float),\n",
+ "})\n",
+ "devices = pd.DataFrame({\n",
+ " \"device_id\": np.arange(N_CUSTOMERS + N_SHARED_DEVICES),\n",
+ " \"is_emulator\": np.r_[np.zeros(N_CUSTOMERS), (rng.random(N_SHARED_DEVICES) < 0.7)].astype(float),\n",
+ "})\n",
+ "recent_customers = customers.customer_id[customers.joined_at > pd.Timestamp(START) - pd.Timedelta(days=180)].to_numpy()\n",
+ "shared_devices = np.arange(N_CUSTOMERS, N_CUSTOMERS + N_SHARED_DEVICES)\n",
+ "\n",
+ "\n",
+ "def simulate_transactions(owners, anchors, owner_devices, window_days, pattern):\n",
+ " n = len(owners)\n",
+ " if pattern == \"everyday\":\n",
+ " amount, intl, hour, merchant = rng.lognormal(3.5, 0.9, n), rng.random(n) < 0.05, rng.normal(14, 4, n), rng.choice(everyday_merchants, n)\n",
+ " device = np.where(rng.random(n) < 0.02, rng.choice(shared_devices, n), owner_devices)\n",
+ " elif pattern == \"lookalike\":\n",
+ " amount = np.where(rng.random(n) < 0.5, rng.lognormal(3.0, 0.6, n), rng.lognormal(5.8, 0.6, n))\n",
+ " intl, hour = rng.random(n) < 0.5, np.where(rng.random(n) < 0.5, rng.normal(2, 2, n), rng.normal(14, 4, n))\n",
+ " merchant = np.where(rng.random(n) < 0.3, rng.choice(np.r_[risky_merchants[\"early\"], risky_merchants[\"late\"]], n), rng.choice(everyday_merchants, n))\n",
+ " device = np.where(rng.random(n) < 0.25, rng.choice(shared_devices, n), owner_devices)\n",
+ " else:\n",
+ " late = pattern == \"late\"\n",
+ " amount = rng.lognormal(3.0, 0.6, n) if late else rng.lognormal(6.0, 0.6, n)\n",
+ " intl, hour = rng.random(n) < (0.6 if late else 0.15), rng.normal(2, 2, n) if late else rng.normal(14, 4, n)\n",
+ " merchant = np.where(rng.random(n) < 0.35, rng.choice(risky_merchants[pattern], n), rng.choice(everyday_merchants, n))\n",
+ " device = np.where(rng.random(n) < (0.35 if late else 0.1), rng.choice(shared_devices, n), owner_devices)\n",
+ " return pd.DataFrame({\n",
+ " \"account_id\": owners, \"merchant_id\": merchant, \"device_id\": device,\n",
+ " \"timestamp\": anchors - pd.to_timedelta(rng.uniform(0, window_days, n), unit=\"D\"),\n",
+ " \"amount\": amount.round(2), \"is_international\": intl.astype(float), \"hour_of_day\": hour.clip(0, 23).round(),\n",
+ " })\n",
+ "\n",
+ "\n",
+ "def simulate_accounts(account_ids, anchors):\n",
+ " n = len(account_ids)\n",
+ " is_fraud = rng.random(n) < FRAUD_RATE\n",
+ " owner = np.where(is_fraud & (rng.random(n) < 0.4), rng.choice(recent_customers, n), rng.integers(0, N_CUSTOMERS, n))\n",
+ " late = (anchors - pd.Timestamp(START)).days.to_numpy() >= DRIFT_DAY\n",
+ " account_frame = pd.DataFrame({\n",
+ " \"account_id\": account_ids, \"customer_id\": owner,\n",
+ " \"opened_at\": np.maximum(\n",
+ " anchors - pd.to_timedelta(np.where(is_fraud, rng.gamma(2.0, 150, n), rng.gamma(3.0, 300, n)).round() + HISTORY_DAYS, unit=\"D\"),\n",
+ " customers.joined_at.to_numpy()[owner] + pd.Timedelta(days=1),\n",
+ " ),\n",
+ " \"credit_limit\": rng.lognormal(8.0, 0.6, n).round(-2),\n",
+ " })\n",
+ " counts = rng.poisson(24, n) + 4\n",
+ " parts = [simulate_transactions(np.repeat(account_ids, counts), np.repeat(anchors.values, counts), np.repeat(owner, counts), HISTORY_DAYS, \"everyday\")]\n",
+ " bursty_fraud = is_fraud & (rng.random(n) < 0.8)\n",
+ " lookalikes = ~is_fraud & (rng.random(n) < 0.04)\n",
+ " for selected, mean_count, pattern in (\n",
+ " (bursty_fraud & ~late, 4, \"early\"), (bursty_fraud & late, 4, \"late\"), (lookalikes, 4, \"lookalike\"),\n",
+ " ):\n",
+ " k = rng.poisson(mean_count, selected.sum()) + 1\n",
+ " parts.append(simulate_transactions(\n",
+ " np.repeat(account_ids[selected], k), np.repeat(anchors.values[selected], k), np.repeat(owner[selected], k), 5, pattern\n",
+ " ))\n",
+ " transaction_frame = pd.concat(parts, ignore_index=True).sort_values(\"timestamp\", ignore_index=True)\n",
+ " transaction_frame[\"timestamp\"] = transaction_frame.timestamp.dt.floor(\"s\")\n",
+ " label_frame = pd.DataFrame({\"account_id\": account_ids, \"timestamp\": anchors, \"is_fraud\": np.where(is_fraud, \"fraud\", \"legit\")})\n",
+ " return account_frame, transaction_frame, label_frame\n",
+ "\n",
+ "\n",
+ "history_anchors = pd.Timestamp(START) + pd.to_timedelta(rng.integers(60, 365, N_ACCOUNTS), unit=\"D\")\n",
+ "accounts, transactions, labels = simulate_accounts(np.arange(N_ACCOUNTS), history_anchors)\n",
+ "transactions.insert(0, \"transaction_id\", np.arange(len(transactions)))\n",
+ "\n",
+ "display(pd.Series({\"customers\": len(customers), \"accounts\": len(accounts), \"transactions\": len(transactions),\n",
+ " \"merchants\": len(merchants), \"devices\": len(devices)}, name=\"rows\").to_frame())\n",
+ "print(f\"Fraud rate: {labels.is_fraud.eq('fraud').mean():.2%} (about 1 in {round(1 / labels.is_fraud.eq('fraud').mean())})\")"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "7b99e336",
+ "metadata": {},
+ "source": [
+ "## Build the relational data\n",
+ "\n",
+ "Declare column types and foreign keys. KumoRelational uses numerical and datetime features. The sampler only uses transactions, accounts and customers dated before each observation date, so another account of the same customer opened later is excluded; merchant and device attributes are static. See [RelationalData](https://nvidia.github.io/structured-data-models/api/generated/sdm.relational.RelationalData.html)."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 5,
+ "id": "fb9900ab",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:48:33.350152Z",
+ "iopub.status.busy": "2026-11-19T01:48:33.350101Z",
+ "iopub.status.idle": "2026-11-19T01:48:34.634192Z",
+ "shell.execute_reply": "2026-11-19T01:48:34.633708Z"
+ }
+ },
+ "outputs": [],
+ "source": [
+ "STYPES = {\n",
+ " \"customers\": {\"customer_id\": \"id\", \"customer_age_years\": \"numerical\", \"joined_at\": \"datetime\", \"kyc_tier\": \"numerical\"},\n",
+ " \"accounts\": {\"account_id\": \"id\", \"customer_id\": \"id\", \"opened_at\": \"datetime\", \"credit_limit\": \"numerical\"},\n",
+ " \"merchants\": {\"merchant_id\": \"id\", \"risk_score\": \"numerical\", \"is_online\": \"numerical\", \"avg_ticket\": \"numerical\"},\n",
+ " \"devices\": {\"device_id\": \"id\", \"is_emulator\": \"numerical\"},\n",
+ " \"transactions\": {\n",
+ " \"transaction_id\": \"id\", \"account_id\": \"id\", \"merchant_id\": \"id\", \"device_id\": \"id\", \"timestamp\": \"datetime\",\n",
+ " \"amount\": \"numerical\", \"is_international\": \"numerical\", \"hour_of_day\": \"numerical\",\n",
+ " },\n",
+ "}\n",
+ "RELATIONSHIPS = [\n",
+ " {\"left_table\": \"accounts\", \"left_column\": \"customer_id\", \"right_table\": \"customers\", \"right_column\": \"customer_id\"},\n",
+ " {\"left_table\": \"transactions\", \"left_column\": \"account_id\", \"right_table\": \"accounts\", \"right_column\": \"account_id\"},\n",
+ " {\"left_table\": \"transactions\", \"left_column\": \"merchant_id\", \"right_table\": \"merchants\", \"right_column\": \"merchant_id\"},\n",
+ " {\"left_table\": \"transactions\", \"left_column\": \"device_id\", \"right_table\": \"devices\", \"right_column\": \"device_id\"},\n",
+ "]\n",
+ "TASK_LINK = {\"task_column\": \"account_id\", \"table\": \"accounts\", \"table_column\": \"account_id\"}\n",
+ "\n",
+ "frames = {\"customers\": customers, \"accounts\": accounts, \"merchants\": merchants, \"devices\": devices, \"transactions\": transactions}\n",
+ "data = sdm.RelationalData(\n",
+ " tables={name: sdm.TableTensor.from_pandas(frame, stypes=STYPES[name]) for name, frame in frames.items()},\n",
+ " relationships=RELATIONSHIPS,\n",
+ ")\n",
+ "sampler = data.sampler(time_columns={\"transactions\": \"timestamp\", \"accounts\": \"opened_at\", \"customers\": \"joined_at\"})\n",
+ "task_table = sdm.TableTensor.from_pandas(labels, stypes={\"account_id\": \"id\", \"timestamp\": \"datetime\", \"is_fraud\": \"categorical\"})"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "792e0cb6",
+ "metadata": {},
+ "source": [
+ "Merchant and device risk sit two hops from an account, through its transactions."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 6,
+ "id": "2b4b84bf",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:48:34.635386Z",
+ "iopub.status.busy": "2026-11-19T01:48:34.635285Z",
+ "iopub.status.idle": "2026-11-19T01:48:34.913391Z",
+ "shell.execute_reply": "2026-11-19T01:48:34.912962Z"
+ }
+ },
+ "outputs": [
+ {
+ "data": {
+ "image/svg+xml": [
+ "\n",
+ "\n",
+ "\n",
+ "\n",
+ "\n"
+ ],
+ "text/plain": [
+ ""
+ ]
+ },
+ "execution_count": 6,
+ "metadata": {},
+ "output_type": "execute_result"
+ }
+ ],
+ "source": [
+ "data.to_graphviz()"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "dbea97e7",
+ "metadata": {},
+ "source": [
+ "## Split accounts by time\n",
+ "\n",
+ "Each label needs 30 days to mature, so each window ends 30 days before the next one starts: context pool before day 220, validation days 250 to 279, test from day 310. Settings are then fixed using only outcomes known before the windows they are applied to."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 7,
+ "id": "483b06b2",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:48:34.914625Z",
+ "iopub.status.busy": "2026-11-19T01:48:34.914541Z",
+ "iopub.status.idle": "2026-11-19T01:48:34.919935Z",
+ "shell.execute_reply": "2026-11-19T01:48:34.919520Z"
+ }
+ },
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " accounts | \n",
+ " fraud_accounts | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | context pool | \n",
+ " 52458 | \n",
+ " 352 | \n",
+ "
\n",
+ " \n",
+ " | validation | \n",
+ " 9814 | \n",
+ " 67 | \n",
+ "
\n",
+ " \n",
+ " | test | \n",
+ " 18022 | \n",
+ " 124 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
"
+ ],
+ "text/plain": [
+ " accounts fraud_accounts\n",
+ "context pool 52458 352\n",
+ "validation 9814 67\n",
+ "test 18022 124"
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ }
+ ],
+ "source": [
+ "observation_day = (labels.timestamp - pd.Timestamp(START)).dt.days.to_numpy()\n",
+ "is_fraud = labels.is_fraud.eq(\"fraud\").to_numpy()\n",
+ "pool_rows = np.flatnonzero(observation_day < VALIDATION_START_DAY - HORIZON_DAYS)\n",
+ "val_rows = np.flatnonzero((observation_day >= VALIDATION_START_DAY) & (observation_day < TEST_START_DAY - HORIZON_DAYS))\n",
+ "test_rows = np.flatnonzero(observation_day >= TEST_START_DAY)\n",
+ "fraud_pool = pool_rows[is_fraud[pool_rows]]\n",
+ "legit_pool = pool_rows[~is_fraud[pool_rows]]\n",
+ "y_val, y_test = is_fraud[val_rows], is_fraud[test_rows]\n",
+ "\n",
+ "display(pd.DataFrame({\n",
+ " \"accounts\": [len(pool_rows), len(val_rows), len(test_rows)],\n",
+ " \"fraud_accounts\": [is_fraud[pool_rows].sum(), y_val.sum(), y_test.sum()],\n",
+ "}, index=[\"context pool\", \"validation\", \"test\"]))"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "d34604c0",
+ "metadata": {},
+ "source": [
+ "## Define the helpers\n",
+ "\n",
+ "`score_contexts` fits the model on each context, scores the queries and averages the probabilities; `score_validation` scores only validation accounts for tuning trials; one context is an ensemble of one. Sampling covers two hops: `num_neighbors` transactions per account, then each transaction's merchant and device, plus up to `SECOND_HOP_NEIGHBORS` accounts per customer. See [RelationalSampler](https://nvidia.github.io/structured-data-models/api/generated/sdm.relational.RelationalSampler.html)."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 8,
+ "id": "34c71ffd",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:48:34.920976Z",
+ "iopub.status.busy": "2026-11-19T01:48:34.920919Z",
+ "iopub.status.idle": "2026-11-19T01:48:35.354644Z",
+ "shell.execute_reply": "2026-11-19T01:48:35.354258Z"
+ }
+ },
+ "outputs": [],
+ "source": [
+ "model = sdm.models.KumoRelational(device=device)\n",
+ "evaluation_query = task_table[torch.as_tensor(np.concatenate([val_rows, test_rows]))].drop_columns(\"is_fraud\")\n",
+ "\n",
+ "\n",
+ "def score_contexts(contexts, query, relational_sampler, num_neighbors=FIRST_HOP_NEIGHBORS, temporal_strategy=\"uniform\", batch_size=10_000):\n",
+ " options = dict(task_link=TASK_LINK, num_neighbors=[num_neighbors, SECOND_HOP_NEIGHBORS], task_time_column=\"timestamp\", temporal_strategy=temporal_strategy)\n",
+ " member_probabilities = []\n",
+ " for context_rows in contexts:\n",
+ " torch.manual_seed(SEED)\n",
+ " context, related = relational_sampler(task_table[torch.as_tensor(context_rows)], **options).to(device)\n",
+ " model.fit(x=context.drop_columns(\"is_fraud\"), y=context[\"is_fraud\"], related_tables=related)\n",
+ " batches = []\n",
+ " for query_batch in query.split(batch_size, -2):\n",
+ " out = model.predict(*relational_sampler(query_batch, **options).to(device))\n",
+ " batches.append(out.numerical[:, list(out.columns[sdm.Stype.numerical]).index(\"fraud\")].float().cpu().numpy())\n",
+ " member_probabilities.append(np.concatenate(batches))\n",
+ " return np.mean(member_probabilities, axis=0)\n",
+ "\n",
+ "\n",
+ "def score_evaluation(contexts, **options):\n",
+ " probabilities = score_contexts(contexts, evaluation_query, sampler, **options)\n",
+ " return probabilities[: len(val_rows)], probabilities[len(val_rows):]\n",
+ "\n",
+ "\n",
+ "def score_validation(contexts, **options):\n",
+ " return score_contexts(contexts, evaluation_query[: len(val_rows)], sampler, **options)\n",
+ "\n",
+ "\n",
+ "def metrics(y_true, probability, threshold):\n",
+ " flagged = probability >= threshold\n",
+ " return {\n",
+ " \"average_precision\": average_precision_score(y_true, probability),\n",
+ " \"auroc\": roc_auc_score(y_true, probability),\n",
+ " \"precision\": precision_score(y_true, flagged, zero_division=0),\n",
+ " \"recall\": recall_score(y_true, flagged),\n",
+ " \"f1\": f1_score(y_true, flagged),\n",
+ " \"threshold\": threshold,\n",
+ " }\n",
+ "\n",
+ "\n",
+ "def max_f1_threshold(y_true, probability):\n",
+ " precision, recall, thresholds = precision_recall_curve(y_true, probability)\n",
+ " f1 = 2 * precision[:-1] * recall[:-1] / np.clip(precision[:-1] + recall[:-1], 1e-12, None)\n",
+ " return thresholds[np.argmax(f1)]\n",
+ "\n",
+ "\n",
+ "def compose_context(fraud_rows, legit_rows, n_legit, fraud_repeat=1):\n",
+ " return np.concatenate([np.tile(fraud_rows, fraud_repeat), rng.choice(legit_rows, n_legit, replace=False)])\n",
+ "\n",
+ "\n",
+ "stages = {}"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "82f4fd9a",
+ "metadata": {},
+ "source": [
+ "## Score the baseline\n",
+ "\n",
+ "Use a random context at the natural fraud rate and the default 0.5 cutoff."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 9,
+ "id": "4a25a0a8",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:48:35.355812Z",
+ "iopub.status.busy": "2026-11-19T01:48:35.355749Z",
+ "iopub.status.idle": "2026-11-19T01:49:27.429778Z",
+ "shell.execute_reply": "2026-11-19T01:49:27.429198Z"
+ }
+ },
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "Fraud examples in the baseline context: 44 of 5000\n"
+ ]
+ },
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " average_precision | \n",
+ " auroc | \n",
+ " precision | \n",
+ " recall | \n",
+ " f1 | \n",
+ " threshold | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | Baseline | \n",
+ " 0.238 | \n",
+ " 0.907 | \n",
+ " 0.818 | \n",
+ " 0.073 | \n",
+ " 0.133 | \n",
+ " 0.5 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
"
+ ],
+ "text/plain": [
+ " average_precision auroc precision recall f1 threshold\n",
+ "Baseline 0.238 0.907 0.818 0.073 0.133 0.5"
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ }
+ ],
+ "source": [
+ "baseline_context = rng.choice(pool_rows, BASELINE_CONTEXT_SIZE, replace=False)\n",
+ "baseline_val, baseline_test = score_evaluation([baseline_context])\n",
+ "stages[\"Baseline\"] = (baseline_val, baseline_test)\n",
+ "print(f\"Fraud examples in the baseline context: {is_fraud[baseline_context].sum()} of {BASELINE_CONTEXT_SIZE}\")\n",
+ "display(pd.DataFrame({\"Baseline\": metrics(y_test, baseline_test, DEFAULT_THRESHOLD)}).T.round(3))"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "53437807",
+ "metadata": {},
+ "source": [
+ "With few positives in context, the model is conservative: reasonable precision, low recall.\n",
+ "\n",
+ "## Strategy 1: rebalance the context\n",
+ "\n",
+ "Every candidate contains all fraud accounts in the pool: ratios of 1:1, 1:5 and 1:20, a 10,000-row context, and a 10,000-row context with each fraud account repeated 10 times. Select by validation average precision, which is threshold-free."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 10,
+ "id": "bfe64ce3",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:49:27.431452Z",
+ "iopub.status.busy": "2026-11-19T01:49:27.431351Z",
+ "iopub.status.idle": "2026-11-19T01:54:02.864755Z",
+ "shell.execute_reply": "2026-11-19T01:54:02.862761Z"
+ }
+ },
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " validation_average_precision | \n",
+ " validation_recall_at_0.5 | \n",
+ " validation_precision_at_0.5 | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | 1:1 | \n",
+ " 0.376 | \n",
+ " 0.896 | \n",
+ " 0.048 | \n",
+ "
\n",
+ " \n",
+ " | 1:5 | \n",
+ " 0.363 | \n",
+ " 0.761 | \n",
+ " 0.079 | \n",
+ "
\n",
+ " \n",
+ " | 1:20 | \n",
+ " 0.406 | \n",
+ " 0.478 | \n",
+ " 0.288 | \n",
+ "
\n",
+ " \n",
+ " | All fraud, 10,000 rows | \n",
+ " 0.387 | \n",
+ " 0.582 | \n",
+ " 0.177 | \n",
+ "
\n",
+ " \n",
+ " | Fraud x10, 10,000 rows | \n",
+ " 0.227 | \n",
+ " 0.448 | \n",
+ " 0.164 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
"
+ ],
+ "text/plain": [
+ " validation_average_precision \\\n",
+ "1:1 0.376 \n",
+ "1:5 0.363 \n",
+ "1:20 0.406 \n",
+ "All fraud, 10,000 rows 0.387 \n",
+ "Fraud x10, 10,000 rows 0.227 \n",
+ "\n",
+ " validation_recall_at_0.5 validation_precision_at_0.5 \n",
+ "1:1 0.896 0.048 \n",
+ "1:5 0.761 0.079 \n",
+ "1:20 0.478 0.288 \n",
+ "All fraud, 10,000 rows 0.582 0.177 \n",
+ "Fraud x10, 10,000 rows 0.448 0.164 "
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ },
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "Selected context design: 1:20\n"
+ ]
+ }
+ ],
+ "source": [
+ "n_fraud = len(fraud_pool)\n",
+ "if FRAUD_REPEAT * n_fraud >= LARGE_CONTEXT_SIZE:\n",
+ " raise ValueError(f\"FRAUD_REPEAT={FRAUD_REPEAT} needs {FRAUD_REPEAT * n_fraud:,} fraud rows, more than LARGE_CONTEXT_SIZE={LARGE_CONTEXT_SIZE:,}.\")\n",
+ "context_designs = {\n",
+ " \"1:1\": (1, n_fraud),\n",
+ " \"1:5\": (1, 5 * n_fraud),\n",
+ " \"1:20\": (1, 20 * n_fraud),\n",
+ " f\"All fraud, {LARGE_CONTEXT_SIZE:,} rows\": (1, LARGE_CONTEXT_SIZE - n_fraud),\n",
+ " f\"Fraud x{FRAUD_REPEAT}, {LARGE_CONTEXT_SIZE:,} rows\": (FRAUD_REPEAT, LARGE_CONTEXT_SIZE - FRAUD_REPEAT * n_fraud),\n",
+ "}\n",
+ "design_scores = {}\n",
+ "for name, (fraud_repeat, n_legit) in context_designs.items():\n",
+ " design_val, design_test = score_evaluation([compose_context(fraud_pool, legit_pool, n_legit, fraud_repeat)])\n",
+ " design_scores[name] = (design_val, design_test)\n",
+ "\n",
+ "comparison = pd.DataFrame({\n",
+ " name: {\n",
+ " \"validation_average_precision\": average_precision_score(y_val, design_val),\n",
+ " \"validation_recall_at_0.5\": recall_score(y_val, design_val >= DEFAULT_THRESHOLD),\n",
+ " \"validation_precision_at_0.5\": precision_score(y_val, design_val >= DEFAULT_THRESHOLD, zero_division=0),\n",
+ " }\n",
+ " for name, (design_val, _) in design_scores.items()\n",
+ "}).T\n",
+ "best_design = comparison.validation_average_precision.idxmax()\n",
+ "fraud_repeat, n_legit = context_designs[best_design]\n",
+ "stages[\"+ Rebalanced context\"] = design_scores[best_design]\n",
+ "display(comparison.round(3))\n",
+ "print(\"Selected context design:\", best_design)"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "b041663c",
+ "metadata": {},
+ "source": [
+ "Fewer negatives per positive also raise the probabilities, which moves precision and recall at 0.5; strategy 4 handles this.\n",
+ "\n",
+ "## Strategy 2: ensemble contexts\n",
+ "\n",
+ "Use the same fraud accounts in every context with a different draw of negatives, then average the members. See [ensembling](https://nvidia.github.io/structured-data-models/icl.html)."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 11,
+ "id": "c99df08e",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:54:02.879184Z",
+ "iopub.status.busy": "2026-11-19T01:54:02.879080Z",
+ "iopub.status.idle": "2026-11-19T01:58:56.336802Z",
+ "shell.execute_reply": "2026-11-19T01:58:56.334063Z"
+ }
+ },
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "Validation AP per member: [0.362 0.402 0.402 0.382 0.385]\n",
+ "Validation AP of the ensemble: 0.423\n"
+ ]
+ }
+ ],
+ "source": [
+ "ensemble_contexts = [compose_context(fraud_pool, legit_pool, n_legit, fraud_repeat) for _ in range(NUM_ESTIMATORS)]\n",
+ "member_scores = [score_evaluation([context]) for context in ensemble_contexts]\n",
+ "member_val_ap = [average_precision_score(y_val, member_val) for member_val, _ in member_scores]\n",
+ "ensemble_val = np.mean([member_val for member_val, _ in member_scores], axis=0)\n",
+ "ensemble_test = np.mean([member_test for _, member_test in member_scores], axis=0)\n",
+ "stages[\"+ Context ensemble\"] = (ensemble_val, ensemble_test)\n",
+ "print(\"Validation AP per member:\", np.round(member_val_ap, 3))\n",
+ "print(f\"Validation AP of the ensemble: {average_precision_score(y_val, ensemble_val):.3f}\")"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "32cfafa0",
+ "metadata": {},
+ "source": [
+ "The ensemble mainly reduces variance from the negative draw; it beats the average member, but not always the best one.\n",
+ "\n",
+ "## Strategy 3: tune sampling\n",
+ "\n",
+ "Test each change on the first ensemble member's context rows, so only the sampling setting changes, and keep it only if validation average precision improves:\n",
+ "\n",
+ "* **Recent accounts:** build the context from the pool's last 90 days at the same class ratio. That window holds fewer positives, so this trial changes context size too. Draw both classes from that window, or the observation date leaks the label.\n",
+ "* **Most recent transactions:** `temporal_strategy=\"last\"` instead of a uniform sample.\n",
+ "* **More neighbors:** more first-hop transactions per account."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 12,
+ "id": "46692a91",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:58:56.352314Z",
+ "iopub.status.busy": "2026-11-19T01:58:56.351774Z",
+ "iopub.status.idle": "2026-11-19T02:00:14.894883Z",
+ "shell.execute_reply": "2026-11-19T02:00:14.892862Z"
+ }
+ },
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " change | \n",
+ " context_rows | \n",
+ " fraud_rows | \n",
+ " num_neighbors | \n",
+ " temporal_strategy | \n",
+ " validation_average_precision | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | 0 | \n",
+ " Current settings | \n",
+ " 7392 | \n",
+ " 352 | \n",
+ " 8 | \n",
+ " uniform | \n",
+ " 0.362 | \n",
+ "
\n",
+ " \n",
+ " | 1 | \n",
+ " Recent accounts | \n",
+ " 2877 | \n",
+ " 137 | \n",
+ " 8 | \n",
+ " uniform | \n",
+ " 0.392 | \n",
+ "
\n",
+ " \n",
+ " | 2 | \n",
+ " Most recent transactions | \n",
+ " 2877 | \n",
+ " 137 | \n",
+ " 8 | \n",
+ " last | \n",
+ " 0.679 | \n",
+ "
\n",
+ " \n",
+ " | 3 | \n",
+ " 16 neighbors | \n",
+ " 2877 | \n",
+ " 137 | \n",
+ " 16 | \n",
+ " last | \n",
+ " 0.609 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
"
+ ],
+ "text/plain": [
+ " change context_rows fraud_rows num_neighbors \\\n",
+ "0 Current settings 7392 352 8 \n",
+ "1 Recent accounts 2877 137 8 \n",
+ "2 Most recent transactions 2877 137 8 \n",
+ "3 16 neighbors 2877 137 16 \n",
+ "\n",
+ " temporal_strategy validation_average_precision \n",
+ "0 uniform 0.362 \n",
+ "1 uniform 0.392 \n",
+ "2 last 0.679 \n",
+ "3 last 0.609 "
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ },
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "Selected sampling settings: {'use_recent': True, 'num_neighbors': 8, 'temporal_strategy': 'last'}\n"
+ ]
+ }
+ ],
+ "source": [
+ "recent_rows = pool_rows[observation_day[pool_rows] >= VALIDATION_START_DAY - RECENT_CONTEXT_DAYS]\n",
+ "recent_fraud, recent_legit = recent_rows[is_fraud[recent_rows]], recent_rows[~is_fraud[recent_rows]]\n",
+ "legit_per_fraud = n_legit / n_fraud\n",
+ "\n",
+ "\n",
+ "def recent_context():\n",
+ " return compose_context(recent_fraud, recent_legit, round(legit_per_fraud * len(recent_fraud)), fraud_repeat)\n",
+ "\n",
+ "\n",
+ "trial_contexts = {False: ensemble_contexts[0], True: recent_context()}\n",
+ "trials = []\n",
+ "\n",
+ "\n",
+ "def record(change, use_recent, num_neighbors, temporal_strategy, val_prob):\n",
+ " context = trial_contexts[use_recent]\n",
+ " trials.append({\"change\": change, \"context_rows\": len(context), \"fraud_rows\": int(is_fraud[context].sum()),\n",
+ " \"num_neighbors\": num_neighbors, \"temporal_strategy\": temporal_strategy,\n",
+ " \"validation_average_precision\": average_precision_score(y_val, val_prob)})\n",
+ " return trials[-1][\"validation_average_precision\"]\n",
+ "\n",
+ "\n",
+ "def try_settings(change, use_recent, num_neighbors, temporal_strategy):\n",
+ " val_prob = score_validation([trial_contexts[use_recent]], num_neighbors=num_neighbors, temporal_strategy=temporal_strategy)\n",
+ " return record(change, use_recent, num_neighbors, temporal_strategy, val_prob)\n",
+ "\n",
+ "\n",
+ "default_settings = {\"use_recent\": False, \"num_neighbors\": FIRST_HOP_NEIGHBORS, \"temporal_strategy\": \"uniform\"}\n",
+ "settings = dict(default_settings)\n",
+ "best_ap = record(\"Current settings\", **settings, val_prob=member_scores[0][0])\n",
+ "for change, update in [(\"Recent accounts\", {\"use_recent\": True}), (\"Most recent transactions\", {\"temporal_strategy\": \"last\"})] + [\n",
+ " (f\"{n} neighbors\", {\"num_neighbors\": n}) for n in NEIGHBOR_OPTIONS\n",
+ "]:\n",
+ " trial_ap = try_settings(change, **{**settings, **update})\n",
+ " if trial_ap > best_ap:\n",
+ " settings, best_ap = {**settings, **update}, trial_ap\n",
+ "\n",
+ "display(pd.DataFrame(trials).round(3))\n",
+ "print(\"Selected sampling settings:\", settings)"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "1a9c6baa",
+ "metadata": {},
+ "source": [
+ "Score the ensemble with the selected settings. It keeps the same members unless the recent pool was selected, which the stage name then states."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 13,
+ "id": "afe53b51",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T02:00:14.912762Z",
+ "iopub.status.busy": "2026-11-19T02:00:14.912459Z",
+ "iopub.status.idle": "2026-11-19T02:03:56.628518Z",
+ "shell.execute_reply": "2026-11-19T02:03:56.626128Z"
+ }
+ },
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " validation_average_precision | \n",
+ " validation_recall_at_0.5 | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | Baseline | \n",
+ " 0.213 | \n",
+ " 0.104 | \n",
+ "
\n",
+ " \n",
+ " | + Rebalanced context | \n",
+ " 0.406 | \n",
+ " 0.478 | \n",
+ "
\n",
+ " \n",
+ " | + Context ensemble | \n",
+ " 0.423 | \n",
+ " 0.507 | \n",
+ "
\n",
+ " \n",
+ " | + Sampling settings (recent pool) | \n",
+ " 0.682 | \n",
+ " 0.642 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
"
+ ],
+ "text/plain": [
+ " validation_average_precision \\\n",
+ "Baseline 0.213 \n",
+ "+ Rebalanced context 0.406 \n",
+ "+ Context ensemble 0.423 \n",
+ "+ Sampling settings (recent pool) 0.682 \n",
+ "\n",
+ " validation_recall_at_0.5 \n",
+ "Baseline 0.104 \n",
+ "+ Rebalanced context 0.478 \n",
+ "+ Context ensemble 0.507 \n",
+ "+ Sampling settings (recent pool) 0.642 "
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ }
+ ],
+ "source": [
+ "sampling_options = {\"num_neighbors\": settings[\"num_neighbors\"], \"temporal_strategy\": settings[\"temporal_strategy\"]}\n",
+ "if settings[\"use_recent\"]:\n",
+ " final_stage = \"+ Sampling settings (recent pool)\"\n",
+ " final_contexts = [trial_contexts[True]] + [recent_context() for _ in range(NUM_ESTIMATORS - 1)]\n",
+ "else:\n",
+ " final_stage = \"+ Sampling settings\"\n",
+ " final_contexts = ensemble_contexts\n",
+ "if settings == default_settings:\n",
+ " final_val, final_test = ensemble_val, ensemble_test\n",
+ "else:\n",
+ " final_val, final_test = score_evaluation(final_contexts, **sampling_options)\n",
+ "stages[final_stage] = (final_val, final_test)\n",
+ "display(pd.DataFrame({\n",
+ " \"validation_average_precision\": {name: average_precision_score(y_val, val_prob) for name, (val_prob, _) in stages.items()},\n",
+ " \"validation_recall_at_0.5\": {name: recall_score(y_val, val_prob >= DEFAULT_THRESHOLD) for name, (val_prob, _) in stages.items()},\n",
+ "}).round(3))"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "6c02d0de",
+ "metadata": {},
+ "source": [
+ "Recall at 0.5 can swing here, even to zero, because these settings also change the probability scale. For example, repeated positives sampled with `\"last\"` become identical copies, which can push every probability down.\n",
+ "\n",
+ "## Strategy 4: tune the threshold per configuration\n",
+ "\n",
+ "Choose the max-F1 cutoff on validation data for every stage. Re-tune whenever the context or sampling changes."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 14,
+ "id": "ebd64c4d",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T02:03:56.666515Z",
+ "iopub.status.busy": "2026-11-19T02:03:56.664691Z",
+ "iopub.status.idle": "2026-11-19T02:03:56.729774Z",
+ "shell.execute_reply": "2026-11-19T02:03:56.729418Z"
+ }
+ },
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " max_f1_threshold | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | Baseline | \n",
+ " 0.161 | \n",
+ "
\n",
+ " \n",
+ " | + Rebalanced context | \n",
+ " 0.720 | \n",
+ "
\n",
+ " \n",
+ " | + Context ensemble | \n",
+ " 0.567 | \n",
+ "
\n",
+ " \n",
+ " | + Sampling settings (recent pool) | \n",
+ " 0.563 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
"
+ ],
+ "text/plain": [
+ " max_f1_threshold\n",
+ "Baseline 0.161\n",
+ "+ Rebalanced context 0.720\n",
+ "+ Context ensemble 0.567\n",
+ "+ Sampling settings (recent pool) 0.563"
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ }
+ ],
+ "source": [
+ "tuned_thresholds = {name: max_f1_threshold(y_val, val_prob) for name, (val_prob, _) in stages.items()}\n",
+ "display(pd.Series(tuned_thresholds, name=\"max_f1_threshold\").round(3).to_frame())"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "46849252",
+ "metadata": {},
+ "source": [
+ "The cutoffs vary widely, so one carried over from another configuration can flag far too many or too few accounts.\n",
+ "\n",
+ "For fixed review capacity, an alternative is maximum recall at validation precision of at least `TARGET_PRECISION`. Compare both on the final configuration."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 15,
+ "id": "fb8f055e",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T02:03:56.733302Z",
+ "iopub.status.busy": "2026-11-19T02:03:56.733239Z",
+ "iopub.status.idle": "2026-11-19T02:03:56.832244Z",
+ "shell.execute_reply": "2026-11-19T02:03:56.831881Z"
+ }
+ },
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " average_precision | \n",
+ " auroc | \n",
+ " precision | \n",
+ " recall | \n",
+ " f1 | \n",
+ " threshold | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | Max F1 | \n",
+ " 0.705 | \n",
+ " 0.969 | \n",
+ " 0.845 | \n",
+ " 0.573 | \n",
+ " 0.683 | \n",
+ " 0.563 | \n",
+ "
\n",
+ " \n",
+ " | Precision >= 50% | \n",
+ " 0.705 | \n",
+ " 0.969 | \n",
+ " 0.707 | \n",
+ " 0.661 | \n",
+ " 0.683 | \n",
+ " 0.389 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
"
+ ],
+ "text/plain": [
+ " average_precision auroc precision recall f1 \\\n",
+ "Max F1 0.705 0.969 0.845 0.573 0.683 \n",
+ "Precision >= 50% 0.705 0.969 0.707 0.661 0.683 \n",
+ "\n",
+ " threshold \n",
+ "Max F1 0.563 \n",
+ "Precision >= 50% 0.389 "
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ }
+ ],
+ "source": [
+ "val_precision, val_recall, candidate_thresholds = precision_recall_curve(y_val, final_val)\n",
+ "meets_precision = np.flatnonzero(val_precision[:-1] >= TARGET_PRECISION)\n",
+ "operating_points = {\"Max F1\": metrics(y_test, final_test, tuned_thresholds[final_stage])}\n",
+ "if len(meets_precision):\n",
+ " precision_threshold = candidate_thresholds[meets_precision[np.argmax(val_recall[:-1][meets_precision])]]\n",
+ " operating_points[f\"Precision >= {TARGET_PRECISION:.0%}\"] = metrics(y_test, final_test, precision_threshold)\n",
+ "else:\n",
+ " print(f\"No validation cutoff reaches {TARGET_PRECISION:.0%} precision; showing max F1 only.\")\n",
+ "display(pd.DataFrame(operating_points).T.round(3))"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "165e1dfc",
+ "metadata": {},
+ "source": [
+ "## Compare the lift\n",
+ "\n",
+ "Each row adds one strategy to the row above and, after the first row, uses its own tuned cutoff. All rows use the same test accounts, and the table is saved to the run directory. **Demonstration only.** A validation-tuned cutoff can miss its target on new data, so recheck it before setting policy."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 16,
+ "id": "e66b8256",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T02:03:56.833391Z",
+ "iopub.status.busy": "2026-11-19T02:03:56.833328Z",
+ "iopub.status.idle": "2026-11-19T02:03:56.990710Z",
+ "shell.execute_reply": "2026-11-19T02:03:56.990347Z"
+ }
+ },
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " average_precision | \n",
+ " auroc | \n",
+ " precision | \n",
+ " recall | \n",
+ " f1 | \n",
+ " threshold | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | Baseline, 0.5 cutoff | \n",
+ " 0.238 | \n",
+ " 0.907 | \n",
+ " 0.818 | \n",
+ " 0.073 | \n",
+ " 0.133 | \n",
+ " 0.500 | \n",
+ "
\n",
+ " \n",
+ " | Baseline, tuned cutoff | \n",
+ " 0.238 | \n",
+ " 0.907 | \n",
+ " 0.223 | \n",
+ " 0.355 | \n",
+ " 0.274 | \n",
+ " 0.161 | \n",
+ "
\n",
+ " \n",
+ " | + Rebalanced context | \n",
+ " 0.319 | \n",
+ " 0.929 | \n",
+ " 0.828 | \n",
+ " 0.194 | \n",
+ " 0.314 | \n",
+ " 0.720 | \n",
+ "
\n",
+ " \n",
+ " | + Context ensemble | \n",
+ " 0.394 | \n",
+ " 0.943 | \n",
+ " 0.374 | \n",
+ " 0.444 | \n",
+ " 0.406 | \n",
+ " 0.567 | \n",
+ "
\n",
+ " \n",
+ " | + Sampling settings (recent pool) | \n",
+ " 0.705 | \n",
+ " 0.969 | \n",
+ " 0.845 | \n",
+ " 0.573 | \n",
+ " 0.683 | \n",
+ " 0.563 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
"
+ ],
+ "text/plain": [
+ " average_precision auroc precision \\\n",
+ "Baseline, 0.5 cutoff 0.238 0.907 0.818 \n",
+ "Baseline, tuned cutoff 0.238 0.907 0.223 \n",
+ "+ Rebalanced context 0.319 0.929 0.828 \n",
+ "+ Context ensemble 0.394 0.943 0.374 \n",
+ "+ Sampling settings (recent pool) 0.705 0.969 0.845 \n",
+ "\n",
+ " recall f1 threshold \n",
+ "Baseline, 0.5 cutoff 0.073 0.133 0.500 \n",
+ "Baseline, tuned cutoff 0.355 0.274 0.161 \n",
+ "+ Rebalanced context 0.194 0.314 0.720 \n",
+ "+ Context ensemble 0.444 0.406 0.567 \n",
+ "+ Sampling settings (recent pool) 0.573 0.683 0.563 "
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ },
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "Average precision: 0.238 -> 0.705; recall: 0.073 -> 0.573; F1: 0.133 -> 0.683\n"
+ ]
+ }
+ ],
+ "source": [
+ "lift_rows = {\"Baseline, 0.5 cutoff\": metrics(y_test, baseline_test, DEFAULT_THRESHOLD)}\n",
+ "for name, (_, test_prob) in stages.items():\n",
+ " lift_rows[\"Baseline, tuned cutoff\" if name == \"Baseline\" else name] = metrics(y_test, test_prob, tuned_thresholds[name])\n",
+ "summary = pd.DataFrame(lift_rows).T\n",
+ "summary.to_csv(run_dir / \"lift_summary.csv\")\n",
+ "display(summary.round(3))\n",
+ "baseline_row, final_row = summary.iloc[0], summary.iloc[-1]\n",
+ "print(f\"Average precision: {baseline_row.average_precision:.3f} -> {final_row.average_precision:.3f}; \"\n",
+ " f\"recall: {baseline_row.recall:.3f} -> {final_row.recall:.3f}; F1: {baseline_row.f1:.3f} -> {final_row.f1:.3f}\")"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "c6d1118c",
+ "metadata": {},
+ "source": [
+ "Plot the headline metrics by stage."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 17,
+ "id": "86a4c153",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T02:03:56.992240Z",
+ "iopub.status.busy": "2026-11-19T02:03:56.992168Z",
+ "iopub.status.idle": "2026-11-19T02:03:57.343190Z",
+ "shell.execute_reply": "2026-11-19T02:03:57.342839Z"
+ }
+ },
+ "outputs": [
+ {
+ "data": {
+ "image/png": "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",
+ "text/plain": [
+ ""
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ }
+ ],
+ "source": [
+ "fig, axes = plt.subplots(1, 3, figsize=(13, 3.8), sharey=True)\n",
+ "for axis, (column, title) in zip(axes, [(\"average_precision\", \"Average precision\"), (\"recall\", \"Recall\"), (\"f1\", \"F1\")]):\n",
+ " values = summary[column].to_numpy()[::-1]\n",
+ " axis.barh(summary.index[::-1], values, color=\"#2a78d6\", height=0.6)\n",
+ " for position, value in enumerate(values):\n",
+ " axis.text(value + 0.01, position, f\"{value:.2f}\", va=\"center\", fontsize=9, color=\"#52514e\")\n",
+ " axis.set_xlim(0, 1)\n",
+ " axis.set_title(title, loc=\"left\", fontsize=11)\n",
+ " axis.spines[[\"top\", \"right\"]].set_visible(False)\n",
+ " axis.grid(axis=\"x\", color=\"#e4e3df\", linewidth=0.8)\n",
+ " axis.set_axisbelow(True)\n",
+ "fig.suptitle(\"Test-set lift by strategy (synthetic data, demonstration only)\", x=0.01, ha=\"left\", fontsize=12)\n",
+ "fig.tight_layout()\n",
+ "plt.show()"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "f25e8b86",
+ "metadata": {},
+ "source": [
+ "## Clean up\n",
+ "\n",
+ "Set `CLEANUP = True` to delete this run's directory."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 18,
+ "id": "b6d3ca27",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T02:03:57.345066Z",
+ "iopub.status.busy": "2026-11-19T02:03:57.344993Z",
+ "iopub.status.idle": "2026-11-19T02:03:57.346817Z",
+ "shell.execute_reply": "2026-11-19T02:03:57.346480Z"
+ }
+ },
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "Results retained in /var/folders/d1/l_l86lcn0jb9jls_r4k74tbc0000gn/T/fraud-context-tuning-c36lf3ed.\n"
+ ]
+ }
+ ],
+ "source": [
+ "if CLEANUP:\n",
+ " workspace.cleanup()\n",
+ "print(\"Removed this run's files.\" if CLEANUP else f\"Results retained in {run_dir}.\")"
+ ]
+ }
+ ],
+ "metadata": {
+ "kernelspec": {
+ "display_name": "Python 3",
+ "language": "python",
+ "name": "python3"
+ },
+ "language_info": {
+ "codemirror_mode": {
+ "name": "ipython",
+ "version": 3
+ },
+ "file_extension": ".py",
+ "mimetype": "text/x-python",
+ "name": "python",
+ "nbconvert_exporter": "python",
+ "pygments_lexer": "ipython3",
+ "version": "3.13.13"
+ }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}
diff --git a/examples/kumo-example-notebooks/structured-data-models/relational/cross-industry/notebooks/local_structured_data_models_regression_tuning.ipynb b/examples/kumo-example-notebooks/structured-data-models/relational/cross-industry/notebooks/local_structured_data_models_regression_tuning.ipynb
new file mode 100644
index 000000000..67b97fe27
--- /dev/null
+++ b/examples/kumo-example-notebooks/structured-data-models/relational/cross-industry/notebooks/local_structured_data_models_regression_tuning.ipynb
@@ -0,0 +1,1641 @@
+{
+ "cells": [
+ {
+ "cell_type": "markdown",
+ "id": "1f14766e",
+ "metadata": {},
+ "source": [
+ "# Improve KumoRelational regression accuracy\n",
+ "\n",
+ "KumoRelational, a relational foundation model (RFM), predicts from labeled in-context examples. This notebook demonstrates four levers and measures the lift of each, using a banking example: how much will each account spend in the next 30 days?\n",
+ "\n",
+ "| Strategy | What changes |\n",
+ "| --- | --- |\n",
+ "| 1. Compose the context | Context size, stratification by target band, and oversampling of the high-value tail. |\n",
+ "| 2. Ensemble contexts | Average the quantile predictions of several context draws. |\n",
+ "| 3. Tune sampling | Recent entities, most recent neighbors, neighbor count. |\n",
+ "| 4. Choose the output quantile | Pick the reported quantile on validation data for each configuration. |\n",
+ "\n",
+ "Runs locally with Python 3.12+ and [direct KumoRelational](https://nvidia.github.io/structured-data-models/api/generated/sdm.models.KumoRelational.html), in about 20 minutes on CPU. Data is synthetic; metrics are **demonstration only**."
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "611aaa51",
+ "metadata": {},
+ "source": [
+ "## Install and configure\n",
+ "\n",
+ "Sampling runs on CPU even when the model uses a GPU, so it always needs a [pyg-lib wheel](https://pytorch-geometric.readthedocs.io/en/latest/install/installation.html) matching your PyTorch version; the schema diagram needs [Graphviz](https://graphviz.org/download/). See [installation](https://nvidia.github.io/structured-data-models/install.html)."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 1,
+ "id": "52d2488b",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:19:53.503325Z",
+ "iopub.status.busy": "2026-11-19T01:19:53.503159Z",
+ "iopub.status.idle": "2026-11-19T01:19:55.288108Z",
+ "shell.execute_reply": "2026-11-19T01:19:55.287403Z"
+ }
+ },
+ "outputs": [],
+ "source": [
+ "%pip install structured-data-models pandas pyarrow scikit-learn scipy matplotlib graphviz"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "a60fff01",
+ "metadata": {},
+ "source": [
+ "## Set the parameters\n",
+ "\n",
+ "Accounts are split by observation date into a context pool, a validation window for choosing settings and quantiles, and a test window for reporting. `TOP_SHARE` defines the high-value tail used for tail metrics and stratification."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 2,
+ "id": "20403f80",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:19:55.289522Z",
+ "iopub.status.busy": "2026-11-19T01:19:55.289404Z",
+ "iopub.status.idle": "2026-11-19T01:19:55.292092Z",
+ "shell.execute_reply": "2026-11-19T01:19:55.291455Z"
+ }
+ },
+ "outputs": [],
+ "source": [
+ "SEED = 7\n",
+ "N_ACCOUNTS = 80_000\n",
+ "START = \"2026-01-01\"\n",
+ "HISTORY_DAYS = 90\n",
+ "HORIZON_DAYS = 30\n",
+ "VALIDATION_START_DAY = 250\n",
+ "TEST_START_DAY = 310\n",
+ "\n",
+ "BASELINE_CONTEXT_SIZE = 5_000\n",
+ "LARGE_CONTEXT_SIZE = 10_000\n",
+ "ROWS_PER_BAND = 2_500\n",
+ "TOP_REPEAT = 3\n",
+ "TOP_SHARE = 0.02\n",
+ "NUM_ESTIMATORS = 5\n",
+ "RECENT_CONTEXT_DAYS = 90\n",
+ "FIRST_HOP_NEIGHBORS = 8\n",
+ "SECOND_HOP_NEIGHBORS = 4\n",
+ "NEIGHBOR_OPTIONS = [16]\n",
+ "\n",
+ "QUANTILES = [round(q, 2) for q in [0.05 * i for i in range(1, 20)]]\n",
+ "UNDER_PREDICTION_COST = 3.0\n",
+ "OUTPUT_PARENT = None\n",
+ "CLEANUP = False"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "8f836116",
+ "metadata": {},
+ "source": [
+ "Import the libraries and create an output directory for this run."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 3,
+ "id": "14c39647",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:19:55.293153Z",
+ "iopub.status.busy": "2026-11-19T01:19:55.293065Z",
+ "iopub.status.idle": "2026-11-19T01:20:03.284724Z",
+ "shell.execute_reply": "2026-11-19T01:20:03.284237Z"
+ }
+ },
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "Device: cpu | Results: /var/folders/d1/l_l86lcn0jb9jls_r4k74tbc0000gn/T/spend-context-tuning-_zl0q29z\n"
+ ]
+ }
+ ],
+ "source": [
+ "import tempfile\n",
+ "from pathlib import Path\n",
+ "\n",
+ "import matplotlib.pyplot as plt\n",
+ "import numpy as np\n",
+ "import pandas as pd\n",
+ "import sdm\n",
+ "import torch\n",
+ "from IPython.display import display\n",
+ "from scipy.stats import spearmanr\n",
+ "\n",
+ "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
+ "rng = np.random.default_rng(SEED)\n",
+ "workspace = tempfile.TemporaryDirectory(prefix=\"spend-context-tuning-\", dir=OUTPUT_PARENT, delete=False)\n",
+ "run_dir = Path(workspace.name)\n",
+ "print(\"Device:\", device, \"| Results:\", run_dir)"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "6d17fce3",
+ "metadata": {},
+ "source": [
+ "## Generate synthetic example data\n",
+ "\n",
+ "Run the next cell to create five tables: customers, accounts, transactions, merchants and devices. You can skim it; the workflow starts at **Build the relational data**.\n",
+ "\n",
+ "About a third of accounts go quiet before their observation date and spend nothing in the next 30 days. About 1% are high spenders who shop at premium merchants. Future spend tracks each account's activity in its last 10 days, and high spenders move from in-store to online premium merchants in late May. Accounts carry an open date, so the sampler can exclude accounts opened after each observation date."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 4,
+ "id": "3efd059e",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:20:03.285952Z",
+ "iopub.status.busy": "2026-11-19T01:20:03.285827Z",
+ "iopub.status.idle": "2026-11-19T01:20:03.675092Z",
+ "shell.execute_reply": "2026-11-19T01:20:03.674457Z"
+ }
+ },
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " rows | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | customers | \n",
+ " 60000 | \n",
+ "
\n",
+ " \n",
+ " | accounts | \n",
+ " 80000 | \n",
+ "
\n",
+ " \n",
+ " | transactions | \n",
+ " 2165327 | \n",
+ "
\n",
+ " \n",
+ " | merchants | \n",
+ " 2000 | \n",
+ "
\n",
+ " \n",
+ " | devices | \n",
+ " 60000 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
"
+ ],
+ "text/plain": [
+ " rows\n",
+ "customers 60000\n",
+ "accounts 80000\n",
+ "transactions 2165327\n",
+ "merchants 2000\n",
+ "devices 60000"
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ },
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " count | \n",
+ " mean | \n",
+ " std | \n",
+ " min | \n",
+ " 35% | \n",
+ " 50% | \n",
+ " 90% | \n",
+ " 99% | \n",
+ " max | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | spend_30d | \n",
+ " 80000.0 | \n",
+ " 347.3 | \n",
+ " 1213.0 | \n",
+ " 0.0 | \n",
+ " 0.0 | \n",
+ " 124.9 | \n",
+ " 789.6 | \n",
+ " 3120.9 | \n",
+ " 108651.6 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
"
+ ],
+ "text/plain": [
+ " count mean std min 35% 50% 90% 99% max\n",
+ "spend_30d 80000.0 347.3 1213.0 0.0 0.0 124.9 789.6 3120.9 108651.6"
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ }
+ ],
+ "source": [
+ "N_CUSTOMERS = int(N_ACCOUNTS * 0.75)\n",
+ "N_MERCHANTS = 2_000\n",
+ "DRIFT_DAY = 150\n",
+ "\n",
+ "merchants = pd.DataFrame({\n",
+ " \"merchant_id\": np.arange(N_MERCHANTS),\n",
+ " \"avg_ticket\": rng.lognormal(3.6, 0.7, N_MERCHANTS).round(2),\n",
+ " \"is_online\": (rng.random(N_MERCHANTS) < 0.4).astype(float),\n",
+ "})\n",
+ "premium = (merchants.avg_ticket > merchants.avg_ticket.quantile(0.9)).to_numpy()\n",
+ "premium_merchants = {\n",
+ " \"early\": merchants.merchant_id[premium & (merchants.is_online == 0)].to_numpy(),\n",
+ " \"late\": merchants.merchant_id[premium & (merchants.is_online == 1)].to_numpy(),\n",
+ "}\n",
+ "everyday_merchants = merchants.merchant_id[~premium].to_numpy()\n",
+ "\n",
+ "customers = pd.DataFrame({\n",
+ " \"customer_id\": np.arange(N_CUSTOMERS),\n",
+ " \"customer_age_years\": rng.integers(18, 85, N_CUSTOMERS).astype(float),\n",
+ " \"income_band\": rng.integers(1, 6, N_CUSTOMERS).astype(float),\n",
+ "})\n",
+ "devices = pd.DataFrame({\n",
+ " \"device_id\": np.arange(N_CUSTOMERS),\n",
+ " \"is_mobile\": (rng.random(N_CUSTOMERS) < 0.7).astype(float),\n",
+ "})\n",
+ "\n",
+ "\n",
+ "def simulate_accounts(account_ids, anchors):\n",
+ " n = len(account_ids)\n",
+ " owner = rng.integers(0, N_CUSTOMERS, n)\n",
+ " income = customers.income_band.to_numpy()[owner]\n",
+ " segment = rng.choice([\"dormant\", \"regular\", \"high\"], n, p=[0.35, 0.64, 0.01])\n",
+ " rate = rng.lognormal(-1.3, 0.5, n)\n",
+ " amount_scale = rng.lognormal(3.3, 0.4, n) * (0.7 + 0.15 * income) * np.where(segment == \"high\", 8.0, 1.0)\n",
+ " premium_share = np.where(segment == \"high\", 0.5, 0.05)\n",
+ " momentum = np.where(segment == \"dormant\", 0.0, rng.lognormal(0, 0.6, n))\n",
+ " quiet_days = np.where(segment == \"dormant\", rng.uniform(12, 30, n), 10.0)\n",
+ "\n",
+ " windows = [(rng.poisson(rate * (HISTORY_DAYS - 10)), np.full(n, float(HISTORY_DAYS)), quiet_days),\n",
+ " (rng.poisson(rate * momentum * 10), np.full(n, 10.0), np.zeros(n))]\n",
+ " parts = []\n",
+ " for counts, oldest, newest in windows:\n",
+ " i = np.repeat(np.arange(n), counts)\n",
+ " timestamp = anchors.values[i] - pd.to_timedelta(rng.uniform(newest[i], oldest[i]), unit=\"D\")\n",
+ " after_switch = (timestamp - pd.Timestamp(START)).days.to_numpy() >= DRIFT_DAY\n",
+ " use_premium = rng.random(len(i)) < premium_share[i]\n",
+ " premium_pick = np.where(after_switch, rng.choice(premium_merchants[\"late\"], len(i)), rng.choice(premium_merchants[\"early\"], len(i)))\n",
+ " merchant = np.where(use_premium, premium_pick, rng.choice(everyday_merchants, len(i)))\n",
+ " parts.append(pd.DataFrame({\n",
+ " \"account_id\": account_ids[i], \"merchant_id\": merchant, \"device_id\": owner[i], \"timestamp\": timestamp,\n",
+ " \"amount\": (amount_scale[i] * rng.lognormal(0, 0.6, len(i)) * np.where(use_premium, 3.0, 1.0)).round(2),\n",
+ " \"hour_of_day\": rng.normal(14, 4, len(i)).clip(0, 23).round(),\n",
+ " }))\n",
+ " transaction_frame = pd.concat(parts, ignore_index=True).sort_values(\"timestamp\", ignore_index=True)\n",
+ " transaction_frame[\"timestamp\"] = transaction_frame.timestamp.dt.floor(\"s\")\n",
+ "\n",
+ " future_spend = rng.poisson(rate * momentum * 30) * amount_scale * (1 + 2 * premium_share) * rng.lognormal(0, 0.3, n)\n",
+ " account_frame = pd.DataFrame({\n",
+ " \"account_id\": account_ids, \"customer_id\": owner,\n",
+ " \"opened_at\": anchors - pd.to_timedelta(rng.gamma(3.0, 300, n).round() + HISTORY_DAYS, unit=\"D\"),\n",
+ " \"credit_limit\": (rng.lognormal(8.0, 0.5, n) * (0.6 + 0.2 * income) * np.where(segment == \"high\", 2.0, 1.0)).round(-2),\n",
+ " })\n",
+ " label_frame = pd.DataFrame({\"account_id\": account_ids, \"timestamp\": anchors, \"spend_30d\": future_spend.round(2)})\n",
+ " return account_frame, transaction_frame, label_frame\n",
+ "\n",
+ "\n",
+ "history_anchors = pd.Timestamp(START) + pd.to_timedelta(rng.integers(60, 365, N_ACCOUNTS), unit=\"D\")\n",
+ "accounts, transactions, labels = simulate_accounts(np.arange(N_ACCOUNTS), history_anchors)\n",
+ "transactions.insert(0, \"transaction_id\", np.arange(len(transactions)))\n",
+ "\n",
+ "display(pd.Series({\"customers\": len(customers), \"accounts\": len(accounts), \"transactions\": len(transactions),\n",
+ " \"merchants\": len(merchants), \"devices\": len(devices)}, name=\"rows\").to_frame())\n",
+ "display(labels.spend_30d.describe(percentiles=[0.35, 0.5, 0.9, 0.99]).round(1).to_frame().T)"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "db4e4aa5",
+ "metadata": {},
+ "source": [
+ "## Build the relational data\n",
+ "\n",
+ "Declare column types and foreign keys. KumoRelational uses numerical and datetime features. The sampler only uses transactions and accounts dated before each observation date, so another account of the same customer opened later is excluded; customer, merchant and device attributes are static. See [RelationalData](https://nvidia.github.io/structured-data-models/api/generated/sdm.relational.RelationalData.html)."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 5,
+ "id": "04413e90",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:20:03.676174Z",
+ "iopub.status.busy": "2026-11-19T01:20:03.676103Z",
+ "iopub.status.idle": "2026-11-19T01:20:05.872017Z",
+ "shell.execute_reply": "2026-11-19T01:20:05.871626Z"
+ }
+ },
+ "outputs": [],
+ "source": [
+ "STYPES = {\n",
+ " \"customers\": {\"customer_id\": \"id\", \"customer_age_years\": \"numerical\", \"income_band\": \"numerical\"},\n",
+ " \"accounts\": {\"account_id\": \"id\", \"customer_id\": \"id\", \"opened_at\": \"datetime\", \"credit_limit\": \"numerical\"},\n",
+ " \"merchants\": {\"merchant_id\": \"id\", \"avg_ticket\": \"numerical\", \"is_online\": \"numerical\"},\n",
+ " \"devices\": {\"device_id\": \"id\", \"is_mobile\": \"numerical\"},\n",
+ " \"transactions\": {\n",
+ " \"transaction_id\": \"id\", \"account_id\": \"id\", \"merchant_id\": \"id\", \"device_id\": \"id\",\n",
+ " \"timestamp\": \"datetime\", \"amount\": \"numerical\", \"hour_of_day\": \"numerical\",\n",
+ " },\n",
+ "}\n",
+ "RELATIONSHIPS = [\n",
+ " {\"left_table\": \"accounts\", \"left_column\": \"customer_id\", \"right_table\": \"customers\", \"right_column\": \"customer_id\"},\n",
+ " {\"left_table\": \"transactions\", \"left_column\": \"account_id\", \"right_table\": \"accounts\", \"right_column\": \"account_id\"},\n",
+ " {\"left_table\": \"transactions\", \"left_column\": \"merchant_id\", \"right_table\": \"merchants\", \"right_column\": \"merchant_id\"},\n",
+ " {\"left_table\": \"transactions\", \"left_column\": \"device_id\", \"right_table\": \"devices\", \"right_column\": \"device_id\"},\n",
+ "]\n",
+ "TASK_LINK = {\"task_column\": \"account_id\", \"table\": \"accounts\", \"table_column\": \"account_id\"}\n",
+ "\n",
+ "frames = {\"customers\": customers, \"accounts\": accounts, \"merchants\": merchants, \"devices\": devices, \"transactions\": transactions}\n",
+ "data = sdm.RelationalData(\n",
+ " tables={name: sdm.TableTensor.from_pandas(frame, stypes=STYPES[name]) for name, frame in frames.items()},\n",
+ " relationships=RELATIONSHIPS,\n",
+ ")\n",
+ "sampler = data.sampler(time_columns={\"transactions\": \"timestamp\", \"accounts\": \"opened_at\"})\n",
+ "task_table = sdm.TableTensor.from_pandas(labels, stypes={\"account_id\": \"id\", \"timestamp\": \"datetime\", \"spend_30d\": \"numerical\"})"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "1befa33b",
+ "metadata": {},
+ "source": [
+ "Merchant spend levels sit two hops from an account, through its transactions."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 6,
+ "id": "ed41df02",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:20:05.873429Z",
+ "iopub.status.busy": "2026-11-19T01:20:05.873296Z",
+ "iopub.status.idle": "2026-11-19T01:20:06.340477Z",
+ "shell.execute_reply": "2026-11-19T01:20:06.340011Z"
+ }
+ },
+ "outputs": [
+ {
+ "data": {
+ "image/svg+xml": [
+ "\n",
+ "\n",
+ "\n",
+ "\n",
+ "\n"
+ ],
+ "text/plain": [
+ ""
+ ]
+ },
+ "execution_count": 6,
+ "metadata": {},
+ "output_type": "execute_result"
+ }
+ ],
+ "source": [
+ "data.to_graphviz()"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "84938429",
+ "metadata": {},
+ "source": [
+ "## Split accounts by time\n",
+ "\n",
+ "Each label needs 30 days to mature, so each window ends 30 days before the next one starts: context pool before day 220, validation days 250 to 279, test from day 310. Settings are then fixed using only outcomes known before the windows they are applied to. Spend bands (zero, low, mid, top) come from the pool's distribution."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 7,
+ "id": "eb0d6472",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:20:06.341913Z",
+ "iopub.status.busy": "2026-11-19T01:20:06.341826Z",
+ "iopub.status.idle": "2026-11-19T01:20:06.349020Z",
+ "shell.execute_reply": "2026-11-19T01:20:06.348674Z"
+ }
+ },
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " accounts | \n",
+ " zero_spend_share | \n",
+ " top_band_accounts | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | context pool | \n",
+ " 41927 | \n",
+ " 0.359 | \n",
+ " 839 | \n",
+ "
\n",
+ " \n",
+ " | validation | \n",
+ " 7902 | \n",
+ " 0.364 | \n",
+ " 179 | \n",
+ "
\n",
+ " \n",
+ " | test | \n",
+ " 14322 | \n",
+ " 0.355 | \n",
+ " 288 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
"
+ ],
+ "text/plain": [
+ " accounts zero_spend_share top_band_accounts\n",
+ "context pool 41927 0.359 839\n",
+ "validation 7902 0.364 179\n",
+ "test 14322 0.355 288"
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ }
+ ],
+ "source": [
+ "observation_day = (labels.timestamp - pd.Timestamp(START)).dt.days.to_numpy()\n",
+ "spend = labels.spend_30d.to_numpy()\n",
+ "pool_rows = np.flatnonzero(observation_day < VALIDATION_START_DAY - HORIZON_DAYS)\n",
+ "val_rows = np.flatnonzero((observation_day >= VALIDATION_START_DAY) & (observation_day < TEST_START_DAY - HORIZON_DAYS))\n",
+ "test_rows = np.flatnonzero(observation_day >= TEST_START_DAY)\n",
+ "y_val, y_test = spend[val_rows], spend[test_rows]\n",
+ "\n",
+ "pool_spend = spend[pool_rows]\n",
+ "top_cut = np.quantile(pool_spend, 1 - TOP_SHARE)\n",
+ "band = np.digitize(spend, [1e-9, np.median(pool_spend[pool_spend > 0]), top_cut])\n",
+ "display(pd.DataFrame({\n",
+ " \"accounts\": [len(pool_rows), len(val_rows), len(test_rows)],\n",
+ " \"zero_spend_share\": [np.mean(spend[r] == 0) for r in (pool_rows, val_rows, test_rows)],\n",
+ " \"top_band_accounts\": [np.sum(band[r] == 3) for r in (pool_rows, val_rows, test_rows)],\n",
+ "}, index=[\"context pool\", \"validation\", \"test\"]).round(3))"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "84d906be",
+ "metadata": {},
+ "source": [
+ "## Define the helpers\n",
+ "\n",
+ "`score_contexts` fits the model on each context, scores the queries and averages each quantile across contexts; `score_validation` scores only validation accounts for tuning trials; one context is an ensemble of one. It keeps the `QUANTILES` grid from the model's 999 outputs. Sampling covers two hops: `num_neighbors` transactions per account, then each transaction's merchant and device, plus up to `SECOND_HOP_NEIGHBORS` accounts per customer. See [RelationalSampler](https://nvidia.github.io/structured-data-models/api/generated/sdm.relational.RelationalSampler.html).\n",
+ "\n",
+ "Metrics: overall MAE, MAE on the top band, Spearman correlation, and top-band capture: the share of actual top-band accounts ranked in the predicted top k, where k is the number of top-band accounts. The top-band metrics are most telling when a few accounts drive most of the value."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 8,
+ "id": "3991dfed",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:20:06.350051Z",
+ "iopub.status.busy": "2026-11-19T01:20:06.349986Z",
+ "iopub.status.idle": "2026-11-19T01:20:07.311089Z",
+ "shell.execute_reply": "2026-11-19T01:20:07.310569Z"
+ }
+ },
+ "outputs": [],
+ "source": [
+ "model = sdm.models.KumoRelational(device=device)\n",
+ "evaluation_query = task_table[torch.as_tensor(np.concatenate([val_rows, test_rows]))].drop_columns(\"spend_30d\")\n",
+ "QUANTILE_COLUMNS = [f\"q{round(q * 1000):03d}\" for q in QUANTILES]\n",
+ "\n",
+ "\n",
+ "def predict_grid(contexts, query, num_neighbors=FIRST_HOP_NEIGHBORS, temporal_strategy=\"uniform\", batch_size=5_000):\n",
+ " options = dict(task_link=TASK_LINK, num_neighbors=[num_neighbors, SECOND_HOP_NEIGHBORS],\n",
+ " task_time_column=\"timestamp\", temporal_strategy=temporal_strategy)\n",
+ " members = []\n",
+ " for context_rows in contexts:\n",
+ " torch.manual_seed(SEED)\n",
+ " context, related = sampler(task_table[torch.as_tensor(context_rows)], **options).to(device)\n",
+ " model.fit(x=context.drop_columns(\"spend_30d\"), y=context[\"spend_30d\"], related_tables=related)\n",
+ " batches = []\n",
+ " for query_batch in query.split(batch_size, -2):\n",
+ " out = model.predict(*sampler(query_batch, **options).to(device))\n",
+ " columns = list(out.columns[sdm.Stype.numerical])\n",
+ " batches.append(out.numerical[:, [columns.index(c) for c in QUANTILE_COLUMNS]].float().cpu().numpy())\n",
+ " members.append(np.concatenate(batches))\n",
+ " return np.mean(members, axis=0)\n",
+ "\n",
+ "\n",
+ "def score_contexts(contexts, **options):\n",
+ " grid = predict_grid(contexts, evaluation_query, **options)\n",
+ " return grid[: len(val_rows)], grid[len(val_rows):]\n",
+ "\n",
+ "\n",
+ "def score_validation(contexts, **options):\n",
+ " return predict_grid(contexts, evaluation_query[: len(val_rows)], **options)\n",
+ "\n",
+ "\n",
+ "def capture(y_true, prediction):\n",
+ " in_band = y_true >= top_cut\n",
+ " k = max(1, int(in_band.sum()))\n",
+ " return in_band[np.argsort(-prediction)[:k]].sum() / k\n",
+ "\n",
+ "\n",
+ "def metrics(y_true, grid, quantile):\n",
+ " prediction = grid[:, QUANTILES.index(quantile)]\n",
+ " top = y_true >= top_cut\n",
+ " return {\n",
+ " \"mae\": np.abs(prediction - y_true).mean(),\n",
+ " \"top_band_mae\": np.abs(prediction[top] - y_true[top]).mean(),\n",
+ " \"spearman\": spearmanr(prediction, y_true).statistic,\n",
+ " \"top_band_capture\": capture(y_true, prediction),\n",
+ " \"quantile\": quantile,\n",
+ " }\n",
+ "\n",
+ "\n",
+ "def best_quantile(y_true, grid, under_cost=1.0):\n",
+ " errors = grid - y_true[:, None]\n",
+ " loss = np.where(errors < 0, -under_cost * errors, errors).mean(axis=0)\n",
+ " return QUANTILES[int(np.argmin(loss))]\n",
+ "\n",
+ "\n",
+ "def tuned_val_mae(val_grid):\n",
+ " return metrics(y_val, val_grid, best_quantile(y_val, val_grid))[\"mae\"]\n",
+ "\n",
+ "\n",
+ "stages = {}"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "3660ef5e",
+ "metadata": {},
+ "source": [
+ "## Score the baseline\n",
+ "\n",
+ "Use a random context at the natural target distribution and report the median (q0.50)."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 9,
+ "id": "b423b5c9",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:20:07.312604Z",
+ "iopub.status.busy": "2026-11-19T01:20:07.312530Z",
+ "iopub.status.idle": "2026-11-19T01:21:23.041636Z",
+ "shell.execute_reply": "2026-11-19T01:21:23.038932Z"
+ }
+ },
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "Top-band accounts in the baseline context: 103 of 5000\n"
+ ]
+ },
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " mae | \n",
+ " top_band_mae | \n",
+ " spearman | \n",
+ " top_band_capture | \n",
+ " quantile | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | Baseline, median | \n",
+ " 253.664 | \n",
+ " 2828.418 | \n",
+ " 0.612 | \n",
+ " 0.448 | \n",
+ " 0.5 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
"
+ ],
+ "text/plain": [
+ " mae top_band_mae spearman top_band_capture quantile\n",
+ "Baseline, median 253.664 2828.418 0.612 0.448 0.5"
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ }
+ ],
+ "source": [
+ "baseline_context = rng.choice(pool_rows, BASELINE_CONTEXT_SIZE, replace=False)\n",
+ "stages[\"Baseline\"] = score_contexts([baseline_context])\n",
+ "print(f\"Top-band accounts in the baseline context: {np.sum(band[baseline_context] == 3)} of {BASELINE_CONTEXT_SIZE}\")\n",
+ "display(pd.DataFrame({\"Baseline, median\": metrics(y_test, stages[\"Baseline\"][1], 0.5)}).T.round(3))"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "54bd64fe",
+ "metadata": {},
+ "source": [
+ "## Strategy 1: compose the context\n",
+ "\n",
+ "Compare a larger random context, equal rows per spend band, every top-band account plus random fill, and the top band repeated `TOP_REPEAT` times. Each candidate is judged at its own validation-tuned quantile, since composition shifts the predicted distribution. This is most useful when the target is skewed, as a random context then holds few high-value examples."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 10,
+ "id": "651212df",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:21:23.069934Z",
+ "iopub.status.busy": "2026-11-19T01:21:23.069602Z",
+ "iopub.status.idle": "2026-11-19T01:27:45.969560Z",
+ "shell.execute_reply": "2026-11-19T01:27:45.966066Z"
+ }
+ },
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " tuned_quantile | \n",
+ " validation_mae | \n",
+ " validation_mae_at_median | \n",
+ " validation_top_band_mae | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | Natural, 5,000 rows | \n",
+ " 0.40 | \n",
+ " 263.107 | \n",
+ " 276.051 | \n",
+ " 3894.181 | \n",
+ "
\n",
+ " \n",
+ " | Natural, 10,000 rows | \n",
+ " 0.45 | \n",
+ " 265.194 | \n",
+ " 266.302 | \n",
+ " 3902.342 | \n",
+ "
\n",
+ " \n",
+ " | Stratified by band | \n",
+ " 0.30 | \n",
+ " 268.686 | \n",
+ " 313.568 | \n",
+ " 3777.482 | \n",
+ "
\n",
+ " \n",
+ " | All top + natural | \n",
+ " 0.35 | \n",
+ " 268.023 | \n",
+ " 297.362 | \n",
+ " 3824.827 | \n",
+ "
\n",
+ " \n",
+ " | Top x3 + natural | \n",
+ " 0.25 | \n",
+ " 285.010 | \n",
+ " 383.530 | \n",
+ " 3809.045 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
"
+ ],
+ "text/plain": [
+ " tuned_quantile validation_mae \\\n",
+ "Natural, 5,000 rows 0.40 263.107 \n",
+ "Natural, 10,000 rows 0.45 265.194 \n",
+ "Stratified by band 0.30 268.686 \n",
+ "All top + natural 0.35 268.023 \n",
+ "Top x3 + natural 0.25 285.010 \n",
+ "\n",
+ " validation_mae_at_median validation_top_band_mae \n",
+ "Natural, 5,000 rows 276.051 3894.181 \n",
+ "Natural, 10,000 rows 266.302 3902.342 \n",
+ "Stratified by band 313.568 3777.482 \n",
+ "All top + natural 297.362 3824.827 \n",
+ "Top x3 + natural 383.530 3809.045 "
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ },
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "Selected context design: Natural, 5,000 rows\n"
+ ]
+ }
+ ],
+ "source": [
+ "def natural(rows, size):\n",
+ " return rng.choice(rows, size, replace=False)\n",
+ "\n",
+ "\n",
+ "def stratified(rows):\n",
+ " return np.concatenate([rng.choice(rows[band[rows] == b], min(ROWS_PER_BAND, np.sum(band[rows] == b)), replace=False) for b in range(4)])\n",
+ "\n",
+ "\n",
+ "def top_plus_natural(rows, repeat):\n",
+ " top_rows = rows[band[rows] == 3]\n",
+ " return np.concatenate([np.tile(top_rows, repeat), natural(rows[band[rows] < 3], LARGE_CONTEXT_SIZE - repeat * len(top_rows))])\n",
+ "\n",
+ "\n",
+ "context_designs = {\n",
+ " f\"Natural, {BASELINE_CONTEXT_SIZE:,} rows\": lambda rows: natural(rows, BASELINE_CONTEXT_SIZE),\n",
+ " f\"Natural, {LARGE_CONTEXT_SIZE:,} rows\": lambda rows: natural(rows, LARGE_CONTEXT_SIZE),\n",
+ " \"Stratified by band\": stratified,\n",
+ " \"All top + natural\": lambda rows: top_plus_natural(rows, 1),\n",
+ " f\"Top x{TOP_REPEAT} + natural\": lambda rows: top_plus_natural(rows, TOP_REPEAT),\n",
+ "}\n",
+ "design_scores = {f\"Natural, {BASELINE_CONTEXT_SIZE:,} rows\": stages[\"Baseline\"]}\n",
+ "for name, design in context_designs.items():\n",
+ " if name not in design_scores:\n",
+ " design_scores[name] = score_contexts([design(pool_rows)])\n",
+ "\n",
+ "comparison = pd.DataFrame({\n",
+ " name: {\n",
+ " \"tuned_quantile\": best_quantile(y_val, val_grid),\n",
+ " \"validation_mae\": tuned_val_mae(val_grid),\n",
+ " \"validation_mae_at_median\": metrics(y_val, val_grid, 0.5)[\"mae\"],\n",
+ " \"validation_top_band_mae\": metrics(y_val, val_grid, best_quantile(y_val, val_grid))[\"top_band_mae\"],\n",
+ " }\n",
+ " for name, (val_grid, _) in design_scores.items()\n",
+ "}).T\n",
+ "best_design = comparison.validation_mae.idxmin()\n",
+ "stages[\"+ Context composition\"] = design_scores[best_design]\n",
+ "display(comparison.round(3))\n",
+ "print(\"Selected context design:\", best_design)"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "aabbb0e1",
+ "metadata": {},
+ "source": [
+ "Oversampling the tail raises the whole predicted distribution, so it needs a lower quantile, and its error at the median grows. Whether it helps after re-tuning depends on how well features already identify the tail. If the baseline design wins, this stage matches the tuned baseline.\n",
+ "\n",
+ "## Strategy 2: ensemble contexts\n",
+ "\n",
+ "Draw `NUM_ESTIMATORS` contexts with the selected design and average each quantile across members. See [ensembling](https://nvidia.github.io/structured-data-models/icl.html)."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 11,
+ "id": "be6bc210",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:27:45.994607Z",
+ "iopub.status.busy": "2026-11-19T01:27:45.993914Z",
+ "iopub.status.idle": "2026-11-19T01:33:31.401527Z",
+ "shell.execute_reply": "2026-11-19T01:33:31.397947Z"
+ }
+ },
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "Validation MAE per member: [268.1 268.4 266.4 263. 259.7]\n",
+ "Validation MAE of the ensemble: 251.3\n"
+ ]
+ }
+ ],
+ "source": [
+ "design = context_designs[best_design]\n",
+ "ensemble_contexts = [design(pool_rows) for _ in range(NUM_ESTIMATORS)]\n",
+ "member_scores = [score_contexts([context]) for context in ensemble_contexts]\n",
+ "stages[\"+ Context ensemble\"] = tuple(np.mean([member[i] for member in member_scores], axis=0) for i in range(2))\n",
+ "print(\"Validation MAE per member:\", np.round([tuned_val_mae(val_grid) for val_grid, _ in member_scores], 1))\n",
+ "print(f\"Validation MAE of the ensemble: {tuned_val_mae(stages['+ Context ensemble'][0]):.1f}\")"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "8bb518e8",
+ "metadata": {},
+ "source": [
+ "## Strategy 3: tune sampling\n",
+ "\n",
+ "Test each change on the first ensemble member's context rows, so only the sampling setting changes, and keep it only if validation MAE, at its tuned quantile, improves:\n",
+ "\n",
+ "* **Recent accounts:** apply the selected design to the pool's last 90 days. This can change the context size too.\n",
+ "* **Most recent transactions:** `temporal_strategy=\"last\"` instead of a uniform sample.\n",
+ "* **More neighbors:** more first-hop transactions per account."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 12,
+ "id": "90e94a4b",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:33:31.420847Z",
+ "iopub.status.busy": "2026-11-19T01:33:31.419794Z",
+ "iopub.status.idle": "2026-11-19T01:35:39.773348Z",
+ "shell.execute_reply": "2026-11-19T01:35:39.771539Z"
+ }
+ },
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " change | \n",
+ " context_rows | \n",
+ " num_neighbors | \n",
+ " temporal_strategy | \n",
+ " validation_mae | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | 0 | \n",
+ " Current settings | \n",
+ " 5000 | \n",
+ " 8 | \n",
+ " uniform | \n",
+ " 268.115 | \n",
+ "
\n",
+ " \n",
+ " | 1 | \n",
+ " Recent accounts | \n",
+ " 5000 | \n",
+ " 8 | \n",
+ " uniform | \n",
+ " 270.096 | \n",
+ "
\n",
+ " \n",
+ " | 2 | \n",
+ " Most recent transactions | \n",
+ " 5000 | \n",
+ " 8 | \n",
+ " last | \n",
+ " 194.881 | \n",
+ "
\n",
+ " \n",
+ " | 3 | \n",
+ " 16 neighbors | \n",
+ " 5000 | \n",
+ " 16 | \n",
+ " last | \n",
+ " 193.014 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
"
+ ],
+ "text/plain": [
+ " change context_rows num_neighbors temporal_strategy \\\n",
+ "0 Current settings 5000 8 uniform \n",
+ "1 Recent accounts 5000 8 uniform \n",
+ "2 Most recent transactions 5000 8 last \n",
+ "3 16 neighbors 5000 16 last \n",
+ "\n",
+ " validation_mae \n",
+ "0 268.115 \n",
+ "1 270.096 \n",
+ "2 194.881 \n",
+ "3 193.014 "
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ },
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "Selected sampling settings: {'use_recent': False, 'num_neighbors': 16, 'temporal_strategy': 'last'}\n"
+ ]
+ }
+ ],
+ "source": [
+ "recent_rows = pool_rows[observation_day[pool_rows] >= VALIDATION_START_DAY - RECENT_CONTEXT_DAYS]\n",
+ "\n",
+ "\n",
+ "trial_contexts = {False: ensemble_contexts[0], True: design(recent_rows)}\n",
+ "trials = []\n",
+ "\n",
+ "\n",
+ "def record(change, use_recent, num_neighbors, temporal_strategy, val_grid):\n",
+ " trials.append({\"change\": change, \"context_rows\": len(trial_contexts[use_recent]), \"num_neighbors\": num_neighbors,\n",
+ " \"temporal_strategy\": temporal_strategy, \"validation_mae\": tuned_val_mae(val_grid)})\n",
+ " return trials[-1][\"validation_mae\"]\n",
+ "\n",
+ "\n",
+ "def try_settings(change, use_recent, num_neighbors, temporal_strategy):\n",
+ " val_grid = score_validation([trial_contexts[use_recent]], num_neighbors=num_neighbors, temporal_strategy=temporal_strategy)\n",
+ " return record(change, use_recent, num_neighbors, temporal_strategy, val_grid)\n",
+ "\n",
+ "\n",
+ "default_settings = {\"use_recent\": False, \"num_neighbors\": FIRST_HOP_NEIGHBORS, \"temporal_strategy\": \"uniform\"}\n",
+ "settings = dict(default_settings)\n",
+ "best_mae = record(\"Current settings\", **settings, val_grid=member_scores[0][0])\n",
+ "for change, update in [(\"Recent accounts\", {\"use_recent\": True}), (\"Most recent transactions\", {\"temporal_strategy\": \"last\"})] + [\n",
+ " (f\"{n} neighbors\", {\"num_neighbors\": n}) for n in NEIGHBOR_OPTIONS\n",
+ "]:\n",
+ " trial_mae = try_settings(change, **{**settings, **update})\n",
+ " if trial_mae < best_mae:\n",
+ " settings, best_mae = {**settings, **update}, trial_mae\n",
+ "\n",
+ "display(pd.DataFrame(trials).round(3))\n",
+ "print(\"Selected sampling settings:\", settings)"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "0aa43f72",
+ "metadata": {},
+ "source": [
+ "Score the ensemble with the selected settings. It keeps the same members unless the recent pool was selected, which the stage name then states."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 13,
+ "id": "adf28a60",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:35:39.789457Z",
+ "iopub.status.busy": "2026-11-19T01:35:39.789340Z",
+ "iopub.status.idle": "2026-11-19T01:46:49.924051Z",
+ "shell.execute_reply": "2026-11-19T01:46:49.910727Z"
+ }
+ },
+ "outputs": [],
+ "source": [
+ "sampling_options = {\"num_neighbors\": settings[\"num_neighbors\"], \"temporal_strategy\": settings[\"temporal_strategy\"]}\n",
+ "if settings[\"use_recent\"]:\n",
+ " final_stage = \"+ Sampling settings (recent pool)\"\n",
+ " final_contexts = [trial_contexts[True]] + [design(recent_rows) for _ in range(NUM_ESTIMATORS - 1)]\n",
+ "else:\n",
+ " final_stage = \"+ Sampling settings\"\n",
+ " final_contexts = ensemble_contexts\n",
+ "if settings == default_settings:\n",
+ " stages[final_stage] = stages[\"+ Context ensemble\"]\n",
+ "else:\n",
+ " stages[final_stage] = score_contexts(final_contexts, **sampling_options)"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "805cc62c",
+ "metadata": {},
+ "source": [
+ "## Strategy 4: choose the output quantile per configuration\n",
+ "\n",
+ "Pick the quantile with the lowest validation MAE for every stage. Re-pick whenever the context or sampling changes. The choice matters more for skewed targets, where the median and the best quantile can sit far apart."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 14,
+ "id": "ee0f322b",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:46:49.952212Z",
+ "iopub.status.busy": "2026-11-19T01:46:49.950723Z",
+ "iopub.status.idle": "2026-11-19T01:46:50.049557Z",
+ "shell.execute_reply": "2026-11-19T01:46:50.049077Z"
+ }
+ },
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " mae_quantile | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | Baseline | \n",
+ " 0.40 | \n",
+ "
\n",
+ " \n",
+ " | + Context composition | \n",
+ " 0.40 | \n",
+ "
\n",
+ " \n",
+ " | + Context ensemble | \n",
+ " 0.45 | \n",
+ "
\n",
+ " \n",
+ " | + Sampling settings | \n",
+ " 0.50 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
"
+ ],
+ "text/plain": [
+ " mae_quantile\n",
+ "Baseline 0.40\n",
+ "+ Context composition 0.40\n",
+ "+ Context ensemble 0.45\n",
+ "+ Sampling settings 0.50"
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ }
+ ],
+ "source": [
+ "tuned_quantiles = {name: best_quantile(y_val, val_grid) for name, (val_grid, _) in stages.items()}\n",
+ "display(pd.Series(tuned_quantiles, name=\"mae_quantile\").to_frame())"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "ac998c1d",
+ "metadata": {},
+ "source": [
+ "When under-forecasting costs more than over-forecasting, for example in liquidity or limit planning, choose the quantile that minimizes that asymmetric cost instead. Compare both on the final configuration."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 15,
+ "id": "a0226264",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:46:50.060704Z",
+ "iopub.status.busy": "2026-11-19T01:46:50.060621Z",
+ "iopub.status.idle": "2026-11-19T01:46:50.103603Z",
+ "shell.execute_reply": "2026-11-19T01:46:50.103113Z"
+ }
+ },
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " mae | \n",
+ " top_band_mae | \n",
+ " spearman | \n",
+ " top_band_capture | \n",
+ " quantile | \n",
+ " asymmetric_cost | \n",
+ " under_forecast_share | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | Min MAE | \n",
+ " 178.940 | \n",
+ " 2116.340 | \n",
+ " 0.838 | \n",
+ " 0.632 | \n",
+ " 0.50 | \n",
+ " 427.971 | \n",
+ " 0.468 | \n",
+ "
\n",
+ " \n",
+ " | Under-forecast cost x3 | \n",
+ " 218.072 | \n",
+ " 2442.324 | \n",
+ " 0.825 | \n",
+ " 0.625 | \n",
+ " 0.65 | \n",
+ " 366.191 | \n",
+ " 0.261 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
"
+ ],
+ "text/plain": [
+ " mae top_band_mae spearman top_band_capture \\\n",
+ "Min MAE 178.940 2116.340 0.838 0.632 \n",
+ "Under-forecast cost x3 218.072 2442.324 0.825 0.625 \n",
+ "\n",
+ " quantile asymmetric_cost under_forecast_share \n",
+ "Min MAE 0.50 427.971 0.468 \n",
+ "Under-forecast cost x3 0.65 366.191 0.261 "
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ }
+ ],
+ "source": [
+ "final_val, final_test = stages[final_stage]\n",
+ "cost_quantile = best_quantile(y_val, final_val, under_cost=UNDER_PREDICTION_COST)\n",
+ "operating_points = {}\n",
+ "for name, quantile in [(\"Min MAE\", tuned_quantiles[final_stage]), (f\"Under-forecast cost x{UNDER_PREDICTION_COST:g}\", cost_quantile)]:\n",
+ " prediction = final_test[:, QUANTILES.index(quantile)]\n",
+ " errors = prediction - y_test\n",
+ " operating_points[name] = {\n",
+ " **metrics(y_test, final_test, quantile),\n",
+ " \"asymmetric_cost\": np.where(errors < 0, -UNDER_PREDICTION_COST * errors, errors).mean(),\n",
+ " \"under_forecast_share\": np.mean(errors < 0),\n",
+ " }\n",
+ "display(pd.DataFrame(operating_points).T.round(3))"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "775cb401",
+ "metadata": {},
+ "source": [
+ "## Compare the lift\n",
+ "\n",
+ "Each row adds one strategy to the row above and, after the first row, reports its own tuned quantile. All rows use the same test accounts, and the table is saved to the run directory. **Demonstration only.**"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 16,
+ "id": "f6dc0fe0",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:46:50.105901Z",
+ "iopub.status.busy": "2026-11-19T01:46:50.105745Z",
+ "iopub.status.idle": "2026-11-19T01:46:50.193359Z",
+ "shell.execute_reply": "2026-11-19T01:46:50.192641Z"
+ }
+ },
+ "outputs": [
+ {
+ "data": {
+ "text/html": [
+ "\n",
+ "\n",
+ "
\n",
+ " \n",
+ " \n",
+ " | \n",
+ " mae | \n",
+ " top_band_mae | \n",
+ " spearman | \n",
+ " top_band_capture | \n",
+ " quantile | \n",
+ "
\n",
+ " \n",
+ " \n",
+ " \n",
+ " | Baseline, median | \n",
+ " 253.664 | \n",
+ " 2828.418 | \n",
+ " 0.612 | \n",
+ " 0.448 | \n",
+ " 0.50 | \n",
+ "
\n",
+ " \n",
+ " | Baseline, tuned quantile | \n",
+ " 246.187 | \n",
+ " 3017.136 | \n",
+ " 0.627 | \n",
+ " 0.444 | \n",
+ " 0.40 | \n",
+ "
\n",
+ " \n",
+ " | + Context composition | \n",
+ " 246.187 | \n",
+ " 3017.136 | \n",
+ " 0.627 | \n",
+ " 0.444 | \n",
+ " 0.40 | \n",
+ "
\n",
+ " \n",
+ " | + Context ensemble | \n",
+ " 236.840 | \n",
+ " 2949.089 | \n",
+ " 0.748 | \n",
+ " 0.476 | \n",
+ " 0.45 | \n",
+ "
\n",
+ " \n",
+ " | + Sampling settings | \n",
+ " 178.940 | \n",
+ " 2116.340 | \n",
+ " 0.838 | \n",
+ " 0.632 | \n",
+ " 0.50 | \n",
+ "
\n",
+ " \n",
+ "
\n",
+ "
"
+ ],
+ "text/plain": [
+ " mae top_band_mae spearman top_band_capture \\\n",
+ "Baseline, median 253.664 2828.418 0.612 0.448 \n",
+ "Baseline, tuned quantile 246.187 3017.136 0.627 0.444 \n",
+ "+ Context composition 246.187 3017.136 0.627 0.444 \n",
+ "+ Context ensemble 236.840 2949.089 0.748 0.476 \n",
+ "+ Sampling settings 178.940 2116.340 0.838 0.632 \n",
+ "\n",
+ " quantile \n",
+ "Baseline, median 0.50 \n",
+ "Baseline, tuned quantile 0.40 \n",
+ "+ Context composition 0.40 \n",
+ "+ Context ensemble 0.45 \n",
+ "+ Sampling settings 0.50 "
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ },
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "MAE: 253.7 -> 178.9; top-band MAE: 2828.4 -> 2116.3; Spearman: 0.612 -> 0.838\n"
+ ]
+ }
+ ],
+ "source": [
+ "lift_rows = {\"Baseline, median\": metrics(y_test, stages[\"Baseline\"][1], 0.5)}\n",
+ "for name, (_, test_grid) in stages.items():\n",
+ " lift_rows[\"Baseline, tuned quantile\" if name == \"Baseline\" else name] = metrics(y_test, test_grid, tuned_quantiles[name])\n",
+ "summary = pd.DataFrame(lift_rows).T\n",
+ "summary.to_csv(run_dir / \"lift_summary.csv\")\n",
+ "display(summary.round(3))\n",
+ "first, last = summary.iloc[0], summary.iloc[-1]\n",
+ "print(f\"MAE: {first.mae:.1f} -> {last.mae:.1f}; top-band MAE: {first.top_band_mae:.1f} -> {last.top_band_mae:.1f}; \"\n",
+ " f\"Spearman: {first.spearman:.3f} -> {last.spearman:.3f}\")"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "766ead17",
+ "metadata": {},
+ "source": [
+ "Plot the headline metrics by stage."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 17,
+ "id": "3158e49b",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:46:50.196218Z",
+ "iopub.status.busy": "2026-11-19T01:46:50.196065Z",
+ "iopub.status.idle": "2026-11-19T01:46:51.042994Z",
+ "shell.execute_reply": "2026-11-19T01:46:51.042223Z"
+ }
+ },
+ "outputs": [
+ {
+ "data": {
+ "image/png": "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",
+ "text/plain": [
+ ""
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ }
+ ],
+ "source": [
+ "panels = [(\"mae\", \"MAE (lower is better)\"), (\"top_band_mae\", \"Top-band MAE (lower is better)\"), (\"spearman\", \"Spearman correlation\")]\n",
+ "fig, axes = plt.subplots(1, 3, figsize=(13, 3.8), sharey=True)\n",
+ "for axis, (column, title) in zip(axes, panels):\n",
+ " values = summary[column].to_numpy()[::-1]\n",
+ " axis.barh(summary.index[::-1], values, color=\"#2a78d6\", height=0.6)\n",
+ " for position, value in enumerate(values):\n",
+ " axis.text(value * 1.01, position, f\"{value:,.2f}\" if column == \"spearman\" else f\"{value:,.0f}\", va=\"center\", fontsize=9, color=\"#52514e\")\n",
+ " axis.set_xlim(0, max(values) * 1.18 if column != \"spearman\" else 1)\n",
+ " axis.set_title(title, loc=\"left\", fontsize=11)\n",
+ " axis.spines[[\"top\", \"right\"]].set_visible(False)\n",
+ " axis.grid(axis=\"x\", color=\"#e4e3df\", linewidth=0.8)\n",
+ " axis.set_axisbelow(True)\n",
+ "fig.suptitle(\"Test-set lift by strategy (synthetic data, demonstration only)\", x=0.01, ha=\"left\", fontsize=12)\n",
+ "fig.tight_layout()\n",
+ "plt.show()"
+ ]
+ },
+ {
+ "cell_type": "markdown",
+ "id": "95ece53a",
+ "metadata": {},
+ "source": [
+ "## Clean up\n",
+ "\n",
+ "Set `CLEANUP = True` to delete this run's directory."
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 18,
+ "id": "cec08f14",
+ "metadata": {
+ "execution": {
+ "iopub.execute_input": "2026-11-19T01:46:51.045662Z",
+ "iopub.status.busy": "2026-11-19T01:46:51.045456Z",
+ "iopub.status.idle": "2026-11-19T01:46:51.049255Z",
+ "shell.execute_reply": "2026-11-19T01:46:51.048433Z"
+ }
+ },
+ "outputs": [
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "Results retained in /var/folders/d1/l_l86lcn0jb9jls_r4k74tbc0000gn/T/spend-context-tuning-_zl0q29z.\n"
+ ]
+ }
+ ],
+ "source": [
+ "if CLEANUP:\n",
+ " workspace.cleanup()\n",
+ "print(\"Removed this run's files.\" if CLEANUP else f\"Results retained in {run_dir}.\")"
+ ]
+ }
+ ],
+ "metadata": {
+ "kernelspec": {
+ "display_name": "Python 3",
+ "language": "python",
+ "name": "python3"
+ },
+ "language_info": {
+ "codemirror_mode": {
+ "name": "ipython",
+ "version": 3
+ },
+ "file_extension": ".py",
+ "mimetype": "text/x-python",
+ "name": "python",
+ "nbconvert_exporter": "python",
+ "pygments_lexer": "ipython3",
+ "version": "3.13.13"
+ }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}