Trending Feed Scoring Module

This module implements a decayed popularity scoring algorithm for generating a trending hook feed. It combines engagement signals with a logarithmic time decay, enabling the system to surface recently engaging content while naturally downranking older items.

Note: This scoring logic is distinct from MPAD's interest-based feed generation. It focuses purely on global popularity without personalization or diversification.

Overview

The scoring formula weighs a hook’s:

Engagement Metrics (views, likes, saves, shares)
Age (in days since creation)

The final score is a decayed sum of weighted interactions, favoring content that is both popular and recent.

Scoring Formula

`decayed_score(hook)`

Formula:

score

Where:

: metric count on type i of hook h
: weight of metric $i$ from POPULARITY_WEIGHTS
: age of the hook in days

Popularity Weights

POPULARITY_WEIGHTS = {
    "viewCount": 0.1,
    "likeCount": 0.5,
    "saveCount": 0.8,
    "shareCount": 1.0,
}

These weights reflect the relative value of different user actions:

Views provide weak signal
Likes and saves reflect growing intent
Shares are the strongest indicator of viral interest

Time Decay

The decay factor:

decay

This ensures that content becomes less relevant as it ages, unless sustained by interaction spikes.

Integration in Feed Pipeline

`update_popularity_scores()`

Iterates over all hooks
Computes their decayed_score
Stores the score in hook.metadata.popularity

`get_trending_feed(N)`

Queries all hooks
Sorts by metadata.popularity (descending)
Returns top N hooks

Output Format

Each result hook includes:

{
  "title": "How to build an LLM app",
  "metadata": {
    "viewCount": 500,
    "likeCount": 120,
    "saveCount": 70,
    "shareCount": 50,
    "popularity": 93.12
  }
}

Design Philosophy

This scoring module is:

Unpersonalized: Same for all users
Lightweight: Easily computable with basic metrics
Dynamic: Changes with new interactions over time