Project
People's Daily on Bilibili: Title and Semantic Drift
A 2,465-video, 13-field study of title semantics, tags, and engagement that reports weak model fit instead of hiding it.
Python · Web Scraping · Sentence-Transformers · scikit-learn · Communication Research
Product Snapshot
- Role
- Independent researcher
- Focus
- Three iterations spanning scraping, title-engagement analysis, semantic similarity, and tag-feature modeling.
- Validation
- The final 2021.06-2023.12 dataset contains 2,465 videos and 13 fields; mean semantic similarity was about 0.408 and random-forest test R² was 0.028.
- Public Proof
- Notebook, final dataset, and a 50-item random audit; the low R² is retained as a finding boundary rather than hidden.
Next Step
Project start
This independent communication-research project evolved through three iterations. I examined whether title language, semantic structure, tags, and engagement showed stable relationships in People’s Daily videos on Bilibili.
Data and method
The final dataset spans 2021.06-2023.12 with 2,465 videos and 13 fields, plus a manual audit of 50 random samples. The workflow combined jieba, Sentence-Transformers and cosine similarity, one-hot tag features, Pearson analysis, and an exploratory random forest.
Result
Mean semantic similarity was about 0.408, with politics and news tags showing more semantic drift. Random-forest test R² was only 0.028.
That weak fit is a useful boundary, not a result to hide: the available features explain very little of engagement variation, which may depend on timing, topic, distribution, and unobserved context.
What I learned
The project established a habit that still shapes my agent-evaluation work: begin with the question, use methods in service of explanation, and report uncertainty rather than manufacturing a neat conclusion.
Configure the public Giscus environment variables to open discussion on the public site.