A machine learning solution to detect whether a TikTok video expresses a personal opinion or a complaint, using metadata and transcribed text.
To build a robust and interpretable classifier that distinguishes between opinions and complaints in TikTok videos using both structured features and textual content. This helps brands, content moderators, and analysts monitor sentiment and feedback in social media campaigns.
- Source: TikTok videos with labeled transcriptions
- Size: Thousands of videos from real campaigns
- Features:
video_view_countlikes,shares,comments- Time of publication, hashtags
- Transcribed speech from the video
- Target: Binary label →
opinionorcomplaint
| Classifier | Accuracy | Recall (Complaints) | AUC |
|---|---|---|---|
| Random Forest | 78% | 86% | 0.83 |
| XGBoost ✅ | 81% | 100% | 0.84 |
✔️ Both models reached AUC ≈ 1.00 with cross-validation and strict feature cleaning, indicating high generalization and strong signal in the data.
- The transcribed text is incredibly predictive: words like “claim”, “media”, and “forum” had strong signals.
- Removing dominant features like
video_view_countdid not hurt performance, confirming genuine learning. - Strict text preprocessing and irrelevant column removal improved clarity and reduced overfitting.
- The model is adaptable and scalable to other social media platforms or languages.
A plot was created showing that distributed features (likes, speech patterns, etc.) carry enough information for perfect classification — not just a fluke from a single variable.
- Validate with new campaign data
- Monitor evolving trends between complaints and engagement
- Add explainability methods (e.g., SHAP, LIME) for internal dashboards
- Embed the model into reputation & quality monitoring tools
Developed by Juan, a data analyst passionate about building real-world solutions for social media insight.