| ### Step 1 - Scraping Data | |
| - Utilizes the CivitAI API with cursor-based pagination to fetch image data over a two-year period. | |
| - Saves progress in `cursors.txt` to resume scraping from the last retrieved point, avoiding redundant requests. | |
| - Stores data in timestamped directories, organizing results into manageable batches of 50,000 images per session. | |
| - Handles API constraints efficiently, with planned improvements for retrying failed requests. | |
| - | |
| # Step 2 - Normalizing Engagement Scores | |
| ### Penalty (Time Penalty) | |
| - **Purpose**: To reduce the influence of older posts by applying a decay based on how long the content has been on the platform. | |
| - **Formula**: \( \text{timePenalty} = \frac{1}{1 + \log(\text{daysOnPlatform} + \text{offset})} \) | |
| - **Logarithmic Decay**: Older posts (higher `daysOnPlatform`) have smaller penalty values. | |
| - **Offset**: Ensures stability and avoids division by zero or undefined log values. | |
| - **Effect**: | |
| - Recent posts get higher scores. | |
| - Older posts are de-emphasized in engagement calculation. | |
| ### Normalizing (Reactions Normalization) | |
| - **Purpose**: To scale raw social reaction values into a standardized range (0–1), enabling comparisons. | |
| - **Method**: Min-Max Scaling | |
| - Formula: \( \text{normalizedReactions} = \frac{\text{value} - \text{min}}{\text{max} - \text{min}} \) | |
| - Adjusts all reaction values relative to the minimum and maximum in the dataset. | |
| - **Effect**: | |
| - Converts raw reaction values into a uniform scale. | |
| - Ensures different ranges of reactions are treated fairly in further calculations. |