RPG-uva

This repository packages two things for publication:

the original RPG implementation, preserved as a read-only dependency in third_party/
the repo-owned reproduction, analysis, and baseline code used to produce the accompanying report

The public baseline name is SASRec. The older duplicate SASRec trees were removed so the repo exposes one baseline surface only.

Repository Structure

third_party/: pinned upstream RPG dependency. Do not edit it here.
scripts/: canonical repo entrypoints. Use scripts/rpg.py and scripts/sasrec.py.
configs/: repo-owned experiment presets. Use configs/rpg/ and configs/sasrec/.
jobs/: Snellius Slurm jobs. Start with the paper index folders: jobs/01_reproduction/, jobs/02_accuracy_and_fairness/, jobs/03_graph_structure_and_dynamics/, jobs/04_search_vs_scorer/, jobs/05_efficiency/.
artifacts/: checkpoints, caches, and runtime outputs.
output/: scheduler stdout/stderr logs.
results/: collected tables and summaries.

Setup

Initialize the submodule:

git submodule update --init --recursive

Create the environment from the repo root:

conda env create -p "$(pwd)/artifacts/conda/rpg-uva" -f environment.yml
conda activate "$(pwd)/artifacts/conda/rpg-uva"

On Snellius, prefer the checked-in Slurm jobs. Submit every job from its own job directory.

RPG's semantic-ID tokenizer needs item-content embeddings. If a dataset's embeddings are not already cached under artifacts/rpg/cache/.../processed/, the tokenizer regenerates them, which by default calls OpenAI's text-embedding-3-large. Before running an RPG job for a new dataset, either:

copy configs/rpg/local.example.yaml to configs/rpg/local.yaml and set openai_api_key, or
override sent_emb_model to a local sentence-transformers encoder instead.

configs/rpg/local.yaml is gitignored, so this file is machine-specific and is not checked in.

Canonical Entry Points

RPG:

python3 scripts/rpg.py --preset beauty

SASRec:

python3 scripts/sasrec.py --preset beauty --dataset Beauty

Reproduction Commands

All commands below call the repo entrypoints directly from the repo root. On Snellius, equivalent checked-in Slurm jobs still live under jobs/. Change datasets as needed.

1. Reproduction

Canonical SASRec full run:

cd /gpfs/home6/$USER/RPG-uva
python3 scripts/sasrec_prepare_data.py --categories Beauty
python3 scripts/sasrec.py --preset beauty --dataset Beauty
python3 scripts/sasrec.py \
  --preset beauty \
  --dataset Beauty \
  --eval-only \
  --checkpoint artifacts/sasrec/ckpt/sasrec_beauty.pt

RPG train and eval:

cd /gpfs/home6/$USER/RPG-uva
python3 scripts/rpg.py --preset beauty

CHECKPOINT_PATH="$(find artifacts/rpg/ckpt -maxdepth 1 -type f -name 'rpg_repro_beauty-*.pth' | sort | tail -n 1)"
python3 scripts/rpg_eval.py \
  --preset beauty \
  --checkpoint "${CHECKPOINT_PATH}" \
  --eval-seed 2024 \
  --num_beams 20 \
  --n_edges 200 \
  --propagation_steps 3

2. Accuracy And Fairness

SASRec multi-seed evaluation:

cd /gpfs/home6/$USER/RPG-uva
python3 scripts/sasrec_eval.py \
  --checkpoint artifacts/sasrec/ckpt/sasrec_sports_and_outdoors.pt \
  --eval-mode eval_seeds \
  --eval-seed 2024 \
  --eval-seeds 2024,2025,2026,2027,2028,2029,2030,2031,2032,2033 \
  --preset sports_and_outdoors \
  --dataset Sports_and_Outdoors \
  --config configs/sasrec/eval_seeds/released_readme/sports_and_outdoors.yaml \
  --output-dir artifacts/sasrec/eval_seeds/released_readme/sports_and_outdoors

RPG multi-seed evaluation:

cd /gpfs/home6/$USER/RPG-uva
CHECKPOINT_PATH="$(find artifacts/rpg/ckpt -maxdepth 1 -type f -name 'rpg_repro_sports_and_outdoors-*.pth' | sort | tail -n 1)"
python3 scripts/rpg_eval_seeds.py \
  --checkpoint "${CHECKPOINT_PATH}" \
  --config configs/rpg/eval_seeds/released_readme/sports_and_outdoors.yaml \
  --eval-seeds 2024,2025,2026,2027,2028,2029,2030,2031,2032,2033 \
  --output-dir artifacts/rpg/eval_seeds/released_readme/sports_and_outdoors \
  --cache_dir artifacts/rpg/cache

SASRec cold-start analysis:

cd /gpfs/home6/$USER/RPG-uva
python3 scripts/sasrec_cold_start.py \
  run \
  --checkpoint artifacts/sasrec/ckpt/sasrec_sports_and_outdoors.pt \
  --preset sports_and_outdoors \
  --dataset Sports_and_Outdoors \
  --output-dir artifacts/sasrec/cold_start

RPG cold-start analysis:

cd /gpfs/home6/$USER/RPG-uva
CHECKPOINT_PATH="$(find artifacts/rpg/ckpt -maxdepth 1 -type f -name 'rpg_repro_sports_and_outdoors-*.pth' | sort | tail -n 1)"
python3 scripts/rpg_cold_start.py \
  run \
  --checkpoint "${CHECKPOINT_PATH}" \
  --config configs/rpg/repro/sports_and_outdoors.yaml \
  --output-dir artifacts/rpg/cold_start \
  --cache_dir artifacts/rpg/cache

3. Graph Structure And Dynamics

Static graph analysis:

cd /gpfs/home6/$USER/RPG-uva
CHECKPOINT_PATH="$(find artifacts/rpg/ckpt -maxdepth 1 -type f -name 'rpg_repro_sports_and_outdoors-*.pth' | sort | tail -n 1)"
SESSION_DIR="artifacts/rpg/graph_analysis/sports/manual_static"
python3 scripts/rpg_graph_analysis.py \
  prepare-graph \
  --checkpoint "${CHECKPOINT_PATH}" \
  --config configs/rpg/graph_analysis/sports.yaml \
  --session-dir "${SESSION_DIR}"
python3 scripts/rpg_graph_analysis.py \
  static \
  --checkpoint "${CHECKPOINT_PATH}" \
  --config configs/rpg/graph_analysis/sports.yaml \
  --session-dir "${SESSION_DIR}"

Dynamic graph analysis:

cd /gpfs/home6/$USER/RPG-uva
CHECKPOINT_PATH="$(find artifacts/rpg/ckpt -maxdepth 1 -type f -name 'rpg_repro_sports_and_outdoors-*.pth' | sort | tail -n 1)"
SESSION_DIR="artifacts/rpg/graph_analysis/sports/manual_static"
python3 scripts/rpg_graph_analysis.py \
  dynamic \
  --checkpoint "${CHECKPOINT_PATH}" \
  --config configs/rpg/graph_analysis/sports.yaml \
  --config configs/rpg/graph_analysis/sports_dynamic.yaml \
  --session-dir "${SESSION_DIR}"

4. Search Vs Scorer

RPG decode grid:

cd /gpfs/home6/$USER/RPG-uva
CHECKPOINT_PATH="$(find artifacts/rpg/ckpt -maxdepth 1 -type f -name 'rpg_sweep_m16_sports_and_outdoors-*.pth' | sort | tail -n 1)"
python3 scripts/rpg_eval_seeds.py \
  --checkpoint "${CHECKPOINT_PATH}" \
  --config configs/rpg/repro/sports_and_outdoors.yaml \
  --split val \
  --eval-seeds 2024,2025,2026 \
  --output-dir output/reproduction/rpg/grid/decode_val/sports_and_outdoors/b20_k200_q3 \
  --n_codebook 16 \
  --num_beams 20 \
  --n_edges 200 \
  --propagation_steps 3 \
  --topk "[5,10]"

RPG decode confirmation:

cd /gpfs/home6/$USER/RPG-uva
CHECKPOINT_PATH="$(find artifacts/rpg/ckpt -maxdepth 1 -type f -name 'rpg_sweep_m16_sports_and_outdoors-*.pth' | sort | tail -n 1)"
python3 scripts/rpg_eval_seeds.py \
  --checkpoint "${CHECKPOINT_PATH}" \
  --config configs/rpg/repro/sports_and_outdoors.yaml \
  --split test \
  --eval-seeds 2024,2025,2026,2027,2028,2029,2030,2031,2032,2033 \
  --output-dir output/reproduction/rpg/grid/decode_test_confirm/sports_and_outdoors/b20_k200_q3 \
  --n_codebook 16 \
  --num_beams 20 \
  --n_edges 200 \
  --propagation_steps 3 \
  --topk "[5,10]" \
  --no-per-user-output

SASRec size ablation:

cd /gpfs/home6/$USER/RPG-uva
python3 scripts/sasrec.py \
  --preset sports_and_outdoors \
  --dataset Sports_and_Outdoors \
  --hidden_size 326 \
  --ckpt_dir artifacts/sasrec/ckpt/ablation_size \
  --run_id sasrec_sports_and_outdoors_size_match

SASRec parameter ablation:

cd /gpfs/home6/$USER/RPG-uva

for lr in 0.001 0.0005 0.0003; do
  python3 scripts/sasrec.py \
    --preset sports_and_outdoors \
    --dataset Sports_and_Outdoors \
    --epochs 300 \
    --lr "${lr}" \
    --hidden_size 326 \
    --ckpt_dir artifacts/sasrec/ckpt/ablation_size/lr_grid \
    --run_id "sasrec_sports_and_outdoors_size_match_e300_lr${lr//./p}"
done

for lr in 0.001 0.0005 0.0003; do
  for layers in 1 2 3; do
    python3 scripts/sasrec.py \
      --preset sports_and_outdoors \
      --dataset Sports_and_Outdoors \
      --epochs 300 \
      --lr "${lr}" \
      --num_hidden_layers "${layers}" \
      --hidden_size 326 \
      --ckpt_dir artifacts/sasrec/ckpt/ablation_size/lr_depth_grid \
      --run_id "sasrec_sports_and_outdoors_size_match_e300_lr${lr//./p}_L${layers}"
  done
done

This reproduces the size-matched SASRec sweeps used for the Search vs Scorer comparison without Slurm: the first loop is the 3-point learning-rate sweep, and the second loop is the 3x3 learning-rate-by-depth sweep (num_hidden_layers in {1,2,3}). For the other paper datasets, keep the same pattern and switch to the matching preset/dataset/hidden size: beauty/Beauty/540, toys_and_games/Toys_and_Games/396, or cds_and_vinyl/CDs_and_Vinyl/328. Checkpoints are written under artifacts/sasrec/ckpt/ablation_size/.

5. Efficiency

RPG inference profiling:

cd /gpfs/home6/$USER/RPG-uva
CHECKPOINT_PATH="$(find artifacts/rpg/ckpt -maxdepth 1 -type f -name 'rpg_repro_sports_and_outdoors-*.pth' | sort | tail -n 1)"
python3 scripts/rpg_perf.py \
  profile \
  --checkpoint "${CHECKPOINT_PATH}" \
  --config configs/rpg/perf/sports.yaml \
  --prepare-only
python3 scripts/rpg_perf.py \
  profile \
  --checkpoint "${CHECKPOINT_PATH}" \
  --config configs/rpg/perf/sports.yaml \
  --profile-only

New-Dataset Extension

The paper datasets live under jobs/reproduction/. Extra datasets such as video_games and pet_supplies live under jobs/new_datasets/.

Use:

cd /gpfs/home6/$USER/RPG-uva/jobs/new_datasets

Then follow jobs/new_datasets/README.md.

Notes

third_party/ stays in the public repository as the preserved original RPG source boundary.
The paper-facing job index lives in jobs/01_reproduction/ through jobs/05_efficiency/.
The canonical SASRec artifacts now live under artifacts/sasrec/ and output/reproduction/sasrec/.

Name		Name	Last commit message	Last commit date
Latest commit History 94 Commits
configs		configs
jobs		jobs
models/sasrec		models/sasrec
notebooks		notebooks
output		output
perf		perf
results		results
rpg_graph_analysis		rpg_graph_analysis
scripts		scripts
third_party @ 7dcf95c		third_party @ 7dcf95c
.gitignore		.gitignore
.gitmodules		.gitmodules
AGENTS.md		AGENTS.md
README.md		README.md
environment.yml		environment.yml
requirements.txt		requirements.txt

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

RPG-uva

Repository Structure

Setup

Canonical Entry Points

Reproduction Commands

1. Reproduction

2. Accuracy And Fairness

3. Graph Structure And Dynamics

4. Search Vs Scorer

5. Efficiency

New-Dataset Extension

Notes

About

Uh oh!

Releases

Packages

Uh oh!

Contributors

Uh oh!

Languages

Folders and files

Latest commit

History

Repository files navigation

RPG-uva

Repository Structure

Setup

Canonical Entry Points

Reproduction Commands

1. Reproduction

2. Accuracy And Fairness

3. Graph Structure And Dynamics

4. Search Vs Scorer

5. Efficiency

New-Dataset Extension

Notes

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages