Skip to content

Feed Reranking Module

This module implements Multifactor Sequential Reranking with Perception-Aware Diversification (MPAD) — a hybrid recommendation engine that combines interest modeling, recency/popularity/exploration reranking, and MMR-based diversification.

Note

This implementation makes specific simplifications and modifications to the original approach proposed in MPAD: Multi-factor Sequential Re-ranking with Perception-Aware Diversification. In particular, we omit the bi-sequential DPP and the perception-aware similarity kernel described in the original model, favoring a more lightweight and interpretable pipeline.

Table of Contents

Overview

The system has three major stages:

  1. User Interest Vector Construction
  2. Hook Scoring with Popularity, Recency, and Exploration Adjustments
  3. Diversified Feed Generation via Maximal Marginal Relevance (MMR)

Key Differences from the Original Paper

This implementation intentionally deviates from the original MPAD model for performance and simplicity:

Collapsible callout

Bi-Sequential Determinantal Point Process (DPP):
Removed in favor of a simpler MMR-based reranking for diversity, improving runtime efficiency and easing tuning.

Collapsible callout

Perception-Aware Kernel for Similarity:
Replaced with a Jaccard similarity on hook tags, allowing interpretable and tag-based diversity computation without complex embedding comparisons.

These changes maintain the core idea of balancing personalization with diversity and exploration, while enabling easier deployment and scaling.

Stage 1: Interest Vector Computation

compute_interest_score(hook, interaction)

Formula:

score formula

  • w_a : weight for action (click, like, etc.) from INTERACTION_WEIGHTS
  • d : interaction duration
  • w_d : weight for duration (from INTERACTION_WEIGHTS['duration'])
  • lambda : decay rate constant (DECAY_LAMBDA)
  • text{daysAgo} : days since hook creation 
score = INTERACTION_WEIGHTS[action] + duration * INTERACTION_WEIGHTS["duration"]
score *= math.exp(-DECAY_LAMBDA * days_ago)

Decay Rationale:

Older content becomes less relevant over time using an exponential decay.

enrich_with_explicit_tags(interest_vector, explicit_tags, implicit_tags)

  • Boosts interest in explicitly and implicitly tagged categories.
  • Boost values:
  • Explicit tag: +5.0
  • Implicit tag: +2.5 (half) 

for tag in explicit_tags:
    interest_vector[tag] += 5.0
for tag in implicit_tags:
    interest_vector[tag] += 2.5

Stage 2: Hook Scoring

compute_reranked_score(hook, base_score, max_views)

Formula:

reranking_score

Where:
- Base score = dot product of interest vector and hook tags
- Recency = e_daysAgo
- Popularity = normalized view count
- Exploration = random factor in [0.01, 0.1]
- Weights abgd from the WEIGHTS dictionary 

score = (
    WEIGHTS["base_score"] * base_score +
    WEIGHTS["recency"] * recency_score +
    WEIGHTS["popularity"] * popularity_score +
    WEIGHTS["exploration_bonus"] * exploration_bonus
)

Purpose:

Balances personal relevance with:
- Recency
- Popularity
- Exploration encouragement

Stage 3: Diversification via MMR

apply_mmr(scored_hooks, N, lambda_param)

MMR Formula:

mmr_formula

  • s : already selected hooks
  • Similarity = Jaccard Similarity on tags:

jacard_similarity

mmr_score = lambda_param * rel_score - (1 - lambda_param) * max_sim

def jaccard(tags1, tags2):
    set1, set2 = set(tags1), set(tags2)
    return len(set1 & set2) / len(set1 | set2)

Diversification Tradeoff:

  • lambda_1 : pure relevance
  • lambda_0 : pure diversity
  • lambda_in_01 : hybrid

Feed Generation Pipeline

update_profile()

  • Iterates all users and logs
  • Computes updated interest vectors
  • Stores them in the profile_collection

get_candidate_hooks(user_interest_vector, N, lambda_param)

  • Collects all hooks
  • Reranks each with compute_reranked_score
  • Sorts and selects top-K candidates (based on CANDIDATE_POOL_FACTOR)
  • Applies MMR to get final top-N diversified results

generate_mpad_feed(user_id, N)

  • Loads user profile
  • Computes feed using reranking and MMR
  • Returns list of hook documents

Constants

These are defined in src.constants:

INTERACTION_WEIGHTS = {
    "clicks": 1.0,
    "likes": 2.0,
    "saves": 2.5,
    "shares": 2.5,
    "duration": 0.05,
}

DECAY_LAMBDA = 0.1

WEIGHTS = {
    "base_score": 0.5,
    "recency": 0.2,
    "popularity": 0.2,
    "exploration_bonus": 0.1,
}

MMR_LAMBDA = 0.7
N_VALUE = 2
CANDIDATE_POOL_FACTOR = 3