josephalexofficial/indabax-kenya-tech4mentalhealth

Multiclass NLP classifier for Kenyan university student mental-health text (Depression, Alcohol, Suicide, Drugs) — IndabaX Kenya / Zindi Tech4MentalHealth hackathon # IndabaX Kenya — Tech4MentalHealth Multiclass NLP system that classifies statements from Kenyan university students into mental-health support categories for a chatbot prototype. Built for the **IndabaX Kenya Tech4MentalHealth Hackathon** on Zindi. --- ## Problem Students respond to the prompt *“What is on your mind?”* with short, noisy, slang-heavy text. The model must assign probabilities for: | Class | Description | |-------|-------------| | **Depression** | Low mood, stress, hopelessness, academic/emotional strain | | **Alcohol** | Alcohol use, craving, quitting, related harm | | **Suicide** | Suicidal ideation, meaning of life, crisis language | | **Drugs** | Substance use (including local terms such as *bhang*) | **Evaluation metric:** Log Loss (lower is better) **Output format:** per-class probabilities in `[0, 1]` (need not sum to 1) --- ## Results (public leaderboard) | # | Submission | Approach | Public Log Loss | |--:|------------|----------|----------------:| | 1 | `01_calibrated_ensemble.csv` | Optimized classical + tree blend with calibration | 0.4173 | | 2 | `02_antioverfit_mix.csv` | Anti-overfit ComplementNB mix with softened prior blend | 0.4409 | | 3 | `03_cnb_svc_blend.csv` | ComplementNB + calibrated LinearSVC (TF-IDF word/char + keywords) | 0.4376 | | 4 | `04_power_from_sub5.csv` | Reverse-psychology upgrade of best lineage (multi-seed bag + pseudo-labels) | 0.4056 | | 5 | `05_last_bullet.csv` | Final upgrade of POWER lineage (7-seed dual-variant + stronger repair) | **0.4002** *(best)* | Lower is better. Best public score achieved: **0.4002**. --- ## Method (final reproducible pipeline) The maintained trainer in `src/train.py` focuses on a strong, interpretable classical stack: 1. **Text normalization** — light spelling fixes while preserving local slang 2. **Features** - TF-IDF word n-grams `(1, 2)` - TF-IDF character n-grams `(3, 5)` (robust to typos) - Domain keyword indicators (depression / alcohol / suicide / drugs) 3. **M …