Skip to content

algorithm-signals.md

Latest commit

 

History

History
85 lines (56 loc) · 3.97 KB

algorithm-signals.md

File metadata and controls

85 lines (56 loc) · 3.97 KB

Signal Processing: Metric Sensitivity & Weighting

Technical Note: The YouTube Shorts recommendation engine operates as a Reinforcement Learning (RL) system. It scores candidate videos based on a predicted reward function, primarily defined by Long-Term Session Value.

This document details the input signals used by the RL agent to calculate the P(satisfaction) score for any given impression.

1. The Hierarchy of Metrics

Metric Hierarchy Pyramid

The algorithm does not weigh all interactions equally. We classify signals into three tiers based on their impact on the Feed Distribution Coefficient (FDC).

Tier 1: The Gatekeepers (Distribution Vetos)

These metrics determine if a video survives the initial seed test (Wave 1).

Metric Technical Name Weight Definition
VVSA Viewed-vs-Swiped-Away 0.65 The percentage of distinct impressions that resulted in a view_complete event (>3s or >10%).
APV Average Percentage Viewed 0.25 The mean completion rate relative to video duration.
ACR Audience Cohort Resonance 0.10 Variance in VVSA between the "Subscriber/History" cohort vs. "Cold/Random" cohort.

Critical Insight: A VVSA below 50% is mathematically impossible to scale in the current model unless APV is >150% (extreme looping).

Tier 2: The Multipliers (Velocity Drivers)

These metrics determine the "Width" of the distribution funnel (Wave 2 & 3).

Metric Technical Name Weight Definition
RVR Re-Watch Rate (Looping) High Implicit signal of high dopamine response or information density. Acts as a multiplier on APV.
Share_Delta Share Velocity Med Total shares normalized by view count. Strong indicator of "external network effects" (Dark Social).
EPV Engagement Per View Low (Likes + Comments) / Total Views. Used primarily for "Safety" checks against clickbait.

Tier 3: Contextual Metadata (Indexing)

These metrics are used for Classification, not Ranking.

  • ASR (Auto-Caption Text): Used for topic extraction (NLP).
  • Visual Object Detection: Frame sampling to detect "Gaming", "Face", "Outdoors".
  • Audio Fingerprint: Music recognition for trending audio clusters.

2. The VVSA Formula

The system calculates a dynamic threshold for VVSA based on video duration.

$$ \text{Threshold}_{eff} = \text{BaseLine} + (\text{Duration}_{sec} \times 0.5) $$

  • 15s Short: Needs ~65% VVSA to break out.
  • 50s Short: Needs ~75% VVSA to break out.

Paradox: Longer Shorts require higher swipe retention because the "opportunity cost" of showing a 50s video is higher for the feed (it blocks 3 potential 15s videos).

3. The "Engagement Trap"

Many creators over-optimize for Likes/Comments. Our data shows a weak correlation between high Like counts and Virality.

Graph: Correlation to View Count (r-value)

  • Loop Rate: r = 0.82 (Strong)
  • Swipe Rate: r = 0.78 (Strong)
  • Share Rate: r = 0.45 (Moderate)
  • Like Rate: r = 0.12 (Weak)
  • Comment Rate: r = 0.09 (Very Weak)

Conclusion: Do not ask for likes. Ask for attention.

4. Signal Decay & Freshness

Signals are not static. The Score(t) of a video decays over time unless refreshed by new interactions.

  • Freshness Boost: New uploads get a distinct multiplier for the first 6 hours.
  • Velocity Decay: A video with 1M views viewing getting 100 views/hour is marked as "Stale" and removed from the active "Viral" queue, moving to "Search/Library" inventory.

The "Resurrection" Trigger

Old videos can re-enter the feed if:

  1. Search Interest Spike: External search volume for the topic rises.
  2. Audio Trend: The BGM used becomes a trending sound.
  3. Cluster Activation: A new viewer cohort (e.g., "People who liked Video Y") is identified as a match for your old Video X.