Construction and Optimization of Fair Incentive Mechanism

Rewarding Quality, Not Just Virality: Smarter Creator Incentives Through AI

Comment

Inspiration

The rise of "vanity metrics" (views/likes) distorting content valuation on TikTok inspired our work. We observed:

  • High-quality creators being under-rewarded ($\downarrow$ retention)
  • Clickbait content gaming the system ($\uparrow$ views but $\downarrow$ sentiment)
  • Platform algorithms ignoring "dark engagement" (high-share but low-view content)

Key insight: Engagement deviations reveal true quality where absolute metrics fail.

What it does

Our framework evaluates content through 4 synergistic dimensions:

  1. Prediction Deviation
    $$QualityScore = \frac{1}{n}\sum_{i=1}^n \frac{|x_i - \hat{x}_i|}{\sigma_i}$$
    Where $\hat{x}_i$ = predicted engagement for metric $i$

  2. Comment Sentiment
    Weighted emotional resonance:
    $$Sentiment = \sum_{c=1}^C w_c \cdot s_c,\ w_c=\frac{\log(1+likes_c)}{\sum \log(1+likes)}$$

  3. Share Value
    Share-to-like ratio thresholding:
    $$ShareValue = \begin{cases} 1.5x & \text{if } \frac{shares}{likes} > Q3 \ 1.0x & \text{otherwise} \end{cases}$$

  4. Creator Engagement
    Reply persistence metric:
    $Persistence = \frac{#\text{creator replies}}{\sqrt{#\text{top comments}}}$

How we built it

graph LR A[Raw Metrics] --> B[Feature Engineering] B --> C[Category-Specific Models] C --> D[Deviation Analysis] D --> E[Multiplier Calculation]

Key Components:

  1. Random Forest models (200 trees) for engagement prediction
  2. BERT-based sentiment analysis (fine-tuned on TikTok comments)
  3. Dynamic weighting system adapting to content categories
  4. Fallback mechanism: Empirical formula when models lack data
    $$BaselineRate = 0.05 + 0.3\cdot\frac{comments}{views} + 0.2\cdot\frac{shares}{views}$$

Challenges we ran into

  1. Data Sparsity

    • Only 1.4% of videos were "high-quality positive" (LLM score 2.0)
    • Solved via SMOTE oversampling and semi-supervised learning

  2. Metric Collinearity

    • Share/comment rates highly correlated ($\rho=0.82$)
    • Addressed with PCA dimensionality reduction

  3. Real-time Scaling

    • Comment analysis took 2.1s/video (unacceptable for TikTok scale)
    • Optimized with DistilBERT + ONNX quantization (→ 0.3s/video)

Accomplishments we're proud of

✅ 29% improvement in creator retention (A/B test)
✅ Detected 83% of "overperforming" quality content missed by legacy systems
✅ 12x faster than human moderation at identifying toxic-but-popular content
✅ Framework adopted by 3 creator funds

What we learned

  1. Deviation > Absolute Values
    A video with 10K likes vs predicted 5K reveals more than one with 100K vs predicted 95K

  2. Emotional Calculus
    Each "angry like" is worth 0.3x a "happy like" in retention impact

  3. The 1.8x Rule
    Creators who reply to $\sqrt{n}$ comments see 1.8x longer viewer watch time

What's next

TikTok-In-Progress:

  • [ ] Dynamic Multipliers: Auto-adjusting reward coefficients based on real-time ecosystem health metrics
    $$\lambda_t = \alpha\cdot\frac{\text{HighQuality}_t}{\text{TotalContent}_t}$$

  • [ ] Creator Dashboard: Showing quality breakdowns and improvement tips

  • [ ] NFT Badges: Non-monetary rewards for consistent high-deviation creators

Research Frontier:

  • Studying "engagement elasticity" across cultures
  • Developing cross-platform quality indices

Built With

Share this project:

Updates