Ranking Signals, Multi-Task Model, and Re-ranking

The Three Signal Categories

Khang Pham's ML Primer organizes feed ranking signals into three groups:

1. Network signals (who posted and your relationship with them):

• User-author affinity: how often has this user liked/replied to/shared this author in the past 90 days? Feature: (likes_from_user_to_author_90d) / (total_impressions_from_author_90d)
• Mutual connection strength: do you have mutual follows? Have you DM'd them?
• Follow recency: newly followed accounts may get a temporary boost (user is actively interested)
• Author credibility: account age, verification status, historical violation rate

2. Content signals (what was posted):

• Topic match: embed post text → cosine similarity with user's interest embedding topic_score = cos_sim(post_topic_emb, user_interest_emb)
• Content type affinity: does this user typically engage more with video vs. text?
• Post quality: grammar, readability, spam signals, link credibility (for link posts)
• Language: post in user's primary language?

3. Engagement signals (how others are reacting):

• Engagement velocity: likes/replies in the first 30 minutes — viral signal
• Global like rate: likes / impressions for this post across all users so far
• Global reply rate: high reply rate = discussion-generating content
• Negative signals: report rate, 'not interested' rate, block rate after seeing this post

Multi-Task Ranking Model

Architecture mirrors content recommendation: shared bottom + task-specific heads.

# Input: concatenation of all signal features
input = concat([
    user_features,       # user embeddings, preferences
    author_features,     # author stats, credibility
    post_features,       # content embedding, type, quality
    user_author_features,# affinity metrics
    engagement_features, # current velocity, rates
    context_features,    # time of day, device, session
])

# Shared bottom → task-specific heads
h = dense_layers(input)
p_like = sigmoid(W_like @ h)        # positive
p_reply = sigmoid(W_reply @ h)      # positive (strong signal)
p_repost = sigmoid(W_repost @ h)    # positive (viral signal)
p_not_interested = sigmoid(W_ni @ h) # negative
p_report = sigmoid(W_report @ h)    # strong negative

score = (1.0*p_like + 2.0*p_reply + 1.5*p_repost
        - 1.0*p_not_interested - 10.0*p_report)

Note: the weights (1.0, 2.0, 1.5, -10.0) are policy decisions, not learned parameters. Policy teams adjust them based on A/B tests and trust-and-safety requirements.

Re-ranking Adjustments

After the ranking model produces scores, apply post-processing:

1. Time-decay: feed_score = ranking_score × 1 / (1 + 0.1 × age_hours) Tune alpha per content type: breaking news decays fast, evergreen content slowly.

2. Author diversity: if the top-10 posts are all from one author (even if they score highest), cap at 3 per author. This prevents single accounts from dominating regardless of score.

3. Deduplication: same story shared by 5 followed accounts → show once. Group by URL hash + text MinHash similarity → keep the version with highest score or earliest original post.

4. Content type mixing: avoid 15 consecutive videos if the user's historical preference is mixed. Interleave content types to match consumption patterns.

5. Seen post filter: remove any post the user already saw in the past 24h (tracked per user in Redis).

6. Ads injection: insert ad units at fixed positions (3, 8, 15) after organic ranking is complete.