Bilibili Trending Analysis System
A full-stack analytics system covering the entire Bilibili trending pipeline: single/multi-day data scraping, content and metric analysis, visualization reports, and multimodal content understanding — all packaged into reusable Skills.
A complete analytics system built around Bilibili’s trending feed — from data scraping and quality analysis to visualization and multimodal understanding, all distilled into reusable, triggerable Skills.
Core Capabilities
The project covers the full workflow: environment health check → data scraping (single-day / multi-day 90-day backfill) → content analysis → data analytics → visualization report → multimodal analysis. It runs in both local and cloud modes, producing a polished report with a single command.
Key Technical Challenges
Reverse-engineering WBI signing. Direct requests to the Bilibili Weekly Must-Watch API return a -352 rate-limit error because the endpoint uses WBI (Web Interface) dynamic signing. By locating already-signed URLs via performance.getEntriesByType('resource') within a loaded page context and reusing them, I bypassed the signing calculation entirely. I also discovered that signatures are bound to episode numbers, requiring an independent page session load per episode.
Full-field data extraction. The opencli adapter only maps 7 fields, while the Bilibili API actually returns 7 engagement metrics (view / danmaku / reply / favorite / coin / share / like) plus metadata like duration, pubdate, and description. I sidestepped the adapter by fetching the raw JSON directly inside the page context.
Analysis Dimensions
A single scraping pass yields ~2,825 unique videos and 2,097 creators. Analysis covers: duration distribution, engagement metric distributions, like-to-view / coin-to-view / favorite-to-view / danmaku-to-view ratios, title features (length / high-frequency words / emoji / punctuation), creator concentration (Lorenz curve), daily trends, ranking decay curves, trending vs. leaderboard radar comparison, trending keyword frequency and engagement correlation matrix. Topic clustering and business insights are generated by an LLM analyzing video titles.
Highlights
- Reverse-engineered Bilibili's WBI signing mechanism to bypass Weekly Must-Watch API rate limiting, enabling 90-day backfill of ~770+ videos
- Single-file visualization report (10+ charts + insight cards) covering duration, engagement rate, title features, creator concentration, and ranking decay
- One scraping pass captures 2,825 unique videos and 2,097 creators, producing topic clustering and business insights
