Skip to content

action-weights.md

Latest commit

 

History

History
224 lines (163 loc) · 8.92 KB

action-weights.md

File metadata and controls

224 lines (163 loc) · 8.92 KB
Raw

Action Weights Reference

Complete reference for all 19 actions the algorithm predicts and weights.

Overview

The X algorithm predicts the probability of 19 different user actions and combines them with weights:

Final Score = Σ (weight_i × P(action_i))

✅ The 19 action heads are confirmed in the downloadable mini Phoenix model config (May 2026 release). The repo's developer notes name 15 of them explicitly: favorite, reply, repost, quote, click, profile click, video view, photo expand, share, dwell, follow author, not interested, block author, mute author, report.

The Exact Weight Values Are Redacted

Verified against home-mixer/scorers/weighted_scorer.rs (May 2026 release). This is the most important thing to understand about "weights":

  • The code multiplies each predicted probability by a per-action weight (p::FAVORITE_WEIGHT, p::REPLY_WEIGHT, p::BLOCK_AUTHOR_WEIGHT, …) and sums them.
  • Those weight values are NOT in the open-source repo. They live in a params module that is not published — home-mixer/lib.rs declares 13 modules and params is not one of them; there is no params.rs and no build script. The code references the constants; the numbers are stripped.

What this means: every specific multiplier in this playbook (e.g. "reply ~2×", "block −10×", "report −20×") is an illustrative estimate / relative ordering — NOT a value extracted from the code. Treat them as direction, not as ground truth. Anyone claiming a precise ratio (e.g. "one reply = 150 likes") is inferring, not quoting the source.

What the source DOES verify

// weighted_scorer.rs — compute_weighted_score() sums exactly 19 terms:
combined =  favorite·FAVORITE_W  + reply·REPLY_W       + retweet·RETWEET_W
          + photo_expand·…       + click·…             + profile_click·…
          + vqv·vqv_weight       + share·…             + share_via_dm·…
          + share_via_copy_link·… + dwell·…            + quote·QUOTE_W
          + quoted_click·…       + dwell_time·CONT_DWELL_TIME_W   // continuous
          + follow_author·…
          + not_interested·NOT_INTERESTED_W            // negative
          + block_author·BLOCK_AUTHOR_W                // negative
          + mute_author·MUTE_AUTHOR_W                  // negative
          + report·REPORT_W;                           // negative

// apply(score, weight) = score.unwrap_or(0.0) * weight  → missing predictions count as 0

Verified facts (no numbers needed):

  • 19 weighted terms, with favorite, reply, retweet, quote, share, share_via_dm, share_via_copy_link, click, quoted_click, profile_click, vqv, photo_expand, dwell, dwell_time, follow_author positive and not_interested, block_author, mute_author, report negative.
  • Video (vqv) only gets weight if video_duration_ms > MIN_VIDEO_DURATION_MS — otherwise its weight is 0.0. (The duration threshold value is also redacted.)
  • dwell_time is continuous (uses CONT_DWELL_TIME_WEIGHT), separate from the binary dwell.
  • After summing, an offset/normalization is applied (offset_score() + normalize_score()): negative-dominant scores are rescaled via NEGATIVE_WEIGHTS_SUM / WEIGHTS_SUM and shifted by NEGATIVE_SCORES_OFFSET, so the final score stays well-behaved.

Positive Actions

Actions that INCREASE your post's score:

# Action Code Name Description Relative Weight
1 Like favorite User likes the post ⭐⭐ Medium
2 Reply reply User replies to post ⭐⭐⭐ High
3 Retweet retweet User retweets ⭐⭐ Medium
4 Quote Tweet quote User quote tweets ⭐⭐⭐ High
5 Share share User shares post ⭐ Low-Medium
6 Share via DM share_via_dm User shares via direct message ⭐ Low
7 Share via Link share_via_copy_link User copies link ⭐ Low
8 Click click User clicks on post ⭐ Low
9 Profile Click profile_click User clicks author's profile ⭐⭐ Medium
10 Follow Author follow_author User follows after seeing post ⭐⭐⭐ High
11 Photo Expand photo_expand User expands photo ⭐ Low
12 Video View video_quality_view User watches video (quality) ⭐⭐ Medium
13 Dwell dwell User stops to read (binary) ⭐ Low
14 Dwell Time dwell_time Time spent reading (continuous) ⭐ Variable
15 Quoted Click quoted_click User clicks quoted content ⭐ Low

Notes on Positive Actions

  • Reply and Quote Tweet are widely treated as the highest-value positive signals (the "~2×" is an estimate, not a code value)
  • Follow Author is high-intent signal (user wants more)
  • Video Quality View only counts if video exceeds minimum duration — verified (vqv_weight = 0 below MIN_VIDEO_DURATION_MS)
  • Dwell Time is continuous (CONT_DWELL_TIME_WEIGHT) — verified

Negative Actions

Actions that DECREASE your post's score:

# Action Code Name Description Relative Weight
16 Not Interested not_interested User clicks "Not interested" ❌ Negative
17 Mute Author mute_author User mutes you ❌❌ Strong Negative
18 Block Author block_author User blocks you ❌❌❌ Very Strong
19 Report report User reports post ❌❌❌❌ Devastating

Negative Weights — Illustrative Only

⚠️ These numbers are NOT from the code (the values are redacted — see above). They're a widely-cited relative intuition for teaching purposes. What the source guarantees is only that these four actions carry negative weight and that report/block are treated as the most severe.

Action Illustrative relative impact Verified from source?
Not Interested ≈ −1× a like Sign only (negative)
Mute ≈ −5× a like Sign only (negative)
Block ≈ −10× a like Sign only (negative)
Report ≈ −20× a like Sign only (negative)

Takeaway (safe to rely on): negative actions actively subtract from your score, and the algorithm separates positive and negative sums with an offset — so a few blocks/reports can erase a lot of positive signal. The exact ratio is unknown.

Weight Categories

Tier 1: Conversation Signals (Highest Value)

reply, quote, follow_author

These indicate deep engagement and intent.

Tier 2: Amplification Signals (High Value)

retweet, share, share_via_dm

User willing to put your content in front of others.

Tier 3: Interest Signals (Medium Value)

favorite, profile_click, video_quality_view

User is interested but passive.

Tier 4: Attention Signals (Low Value)

click, photo_expand, dwell, quoted_click

Basic attention metrics.

Tier -1: Negative Signals (Subtract)

not_interested, mute, block, report

Actively hurts your score.

Special Cases

Video Quality View (VQV)

// Only applies if video exceeds minimum duration
if video_duration_ms > MIN_VIDEO_DURATION_MS {
    apply(vqv_score, VQV_WEIGHT)
} else {
    // VQV weight = 0, not counted
}

Implication: Short video clips don't get the VQV bonus.

Dwell Time (Continuous)

Unlike other binary actions, dwell time is continuous:

dwell_score: Binary (did user stop?)
dwell_time: Continuous (how long in seconds?)

Both are weighted and contribute to score.

Algorithm Source Files

Component Where
Weight application (19-term sum + offset) home-mixer/scorers/weighted_scorer.rs
Score extraction (Phoenix Scorer) home-mixer/scorers/phoenix_scorer.rs
19 action heads phoenix/ model config
Weight VALUES (*_WEIGHT constants) params module — redacted / not published

Strategic Implications

Optimize For (Priority Order)

  1. Replies - Ask questions, create debate
  2. Quote Tweets - Create quotable insights
  3. Follows - Provide unique ongoing value
  4. Retweets - Make shareable content
  5. Likes - Be generally engaging

Avoid At All Costs

  1. Reports - Stay within ToS
  2. Blocks - Don't annoy or spam
  3. Mutes - Don't over-post or be off-brand
  4. "Not Interested" - Stay relevant to audience

Quick Reference Card

MAXIMIZE:
├── Reply (⭐⭐⭐)
├── Quote (⭐⭐⭐)
├── Follow (⭐⭐⭐)
├── Retweet (⭐⭐)
├── Like (⭐⭐)
└── Video View (⭐⭐) [if > min duration]

MINIMIZE (severity order — exact weights redacted):
├── Report (❌❌❌❌ most severe)
├── Block (❌❌❌)
├── Mute (❌❌)
└── Not Interested (❌)

Next: Filter System →