Overview01
Harmful comments reduce user trust, skew engagement metrics, and expose vulnerable viewers to scams and worse. The plan is to classify comments into four buckets, phishing_scam (crypto and messaging-app links), adult_bait (automated content traps), engagement_spam (repetitive “who is watching in 2026” patterns), and benign, so the harmful ones can be surfaced and handled.
The plan02
- Collect. Pull comments through the YouTube Data API.
- Label. Use rule-based weak supervision to flag obvious spam, then validate a sample by hand against clear guidelines.
- Engineer features. Extract behavioral markers: link density, emoji frequency, and repetition signatures.
- Train. Start with a simple text-classification baseline (TF-IDF with logistic regression) for the four categories.
Evaluation plan03
Because a missed scam costs more than a false alarm, the plan is to prioritize recall on the high-risk categories (phishing and adult-bait), while tracking precision so moderators aren’t flooded, and to use confusion matrices and error analysis to show where it works and where it slips.
Tech
PythonYouTube Data APIpandasscikit-learn