GO DS 4.0 - Mental Health Text Classification

This repository contains my solution for the GO DATA SCIENCE 4.0 Hackathon hosted on Zindi, where I achieved 42rd place out of 194 participants. The challenge focused on classifying mental health-related text discussions into predefined categories using Natural Language Processing (NLP) techniques.

---

🏅 Competition Results

Final Leaderboard Performance

• Rank: 42 out of 194 participants
• Validation Accuracy: 74.4%
• Public Leaderboard Score: 0.7528
• Private Leaderboard Score: 0.7371

Top Performers (Excerpt)

Rank	Team Name	Public Score	Private Score
1	Recursive Duo	0.8189	0.7996
...	...	...	...
9	one crew	0.7786	0.7792
25	Llama	0.7610	0.7648
43	SamehAissa (Me)	0.7686	0.7528
44	...	...	...

---

📊 Analysis

Key Observations

1. Top Scores:

   • The winning team achieved a public score of 0.8189 and a private score of 0.7996.

2. My Performance:

   • Achieved a public score of 0.7686 and a private score of 0.7528.
   • Ranked 42rd, placing in the top 22% of participants.

3. Leaderboard Insights:

   • A small gap between public and private scores indicates robust models.
   • The competition was highly competitive, with close scores among top teams.

---

🏆 Competition Overview

Problem Statement

The goal was to develop a model that accurately classifies text entries (titles and content) from online discussions into categories representing mental health issues. Each entry in the dataset included:

• id: Unique identifier
• title: Discussion title
• content: Main body of the text
• target: Mental health category (only in training data)

Example Data Entry

id	title	content	target
101	Feeling Hopeless and Lost	I've been struggling with depression for a while...	Depression
102	Panic Attacks Are Getting Worse	Lately, my panic attacks have been more frequent...	Anxiety

Evaluation Metric

The model's performance was evaluated using Private Accuracy as the primary metric.

---

Key Steps

1. Data Preprocessing:

   • Combined title and content into a single text feature.
   • Handled missing values and cleaned text data.
   • Encoded target labels into numerical format.

2. Modeling:

   • Experimented with BERT and RoBERTa architectures.
   • Implemented class weighting to handle imbalanced data.
   • Used Text Augmentation (EDA) to improve generalization.

3. Training:

   • Fine-tuned transformer models using Hugging Face's Trainer API.
   • Applied Focal Loss to focus on hard-to-classify examples.
   • Used Test-Time Augmentation (TTA) for robust predictions.

4. Evaluation:

   • Achieved ~76.8% Public Accuracy on the validation set.
   • Secured 42rd place on the final leaderboard.

---

🔮 Future Improvements

1. Deployement:

   • Build a Streamlit/Gradio app for real-time predictions.
   • Deploy the model using FastAPI or Flask.

2. Explainability:

   • Use SHAP or LIME to explain model predictions.
   • Visualize attention weights for transformer models.

3. Advanced Models:

   • Experiment with DeBERTa, GPT-based models, or ensemble methods.
   • Use knowledge distillation to combine multiple models.