LAB::SCORE

Weighted Score Lab

Adjust the weights for each action to see how the final score changes. Understand the weighted sum logic in Phoenix Scoring.

Weight Adjustment

Positive Actions

60%
40%
50%
70%
55%
35%

Negative Actions

10%
5%
2%

Presets & Strategy Meanings

Simulates high RT/Reply counts; score grows fastest.
Maximizes Share weight for Out-of-network growth.
Focuses on Likes/Dwell; lowers penalty risks.

Score Calculation

Weighted Formula
Score = Σ (weight × P(action))
FINAL SCORE
0.00
Low
Threshold
Viral

Score Breakdown

Positive Contribution +0.00
Negative Deduction -0.00
Total 0.00
i

Insight

Adjust the weights to see which factors have the biggest impact on the score.

Core Algorithmic Principles

STEP 01

Multi-Target Prediction (Phoenix Scorer)

As a user scrolls, the Phoenix Model (a Transformer-based ranker) predicts a set of engagement probabilities (P(action)) for that specific user. These predictions are **independent and simultaneous**, e.g., "What is the probability this user will Favorite this tweet?" or "What is the probability they will Reply?"

STEP 02

Weighted Summation (Weighted Scorer)

The system does not value all actions equally. Different interactions contribute differently to a tweet's "Heat Score." The Weighted Scorer multiplies each predicted probability by a predefined system weight: 

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

For example, a **Reply** usually carries significantly more weight than a Like. This is why "active conversations" drive much more reach than passive engagement.

STEP 03

Negative Signal Penalties

Critically, the system applies massive **negative weights** to actions like "Not Interested," "Mute," or "Report." Even if P(Like) is high, a slight increase in predicted negative feedback can cause the final score to crash, effectively stopping content distribution.

Back to Tools