Skip to content

When deep learning loses: mental-health text classification on 992 samples

Arzoo Ghani Khan, Rehan Sathio

Status
Completed
Date
2026
Venue
MS AI coursework, Bahria University
Tags
NLP · Text classification · Small data

  • 53%

    Random Forest (TF-IDF)

  • 34%

    Hybrid CNN-BiLSTM

  • 992

    samples

    4 classes

Abstract

A four-class mental-health text classifier (Anxiety, Depression, Normal, Suicidal) trained on a 992-sample Kaggle dataset. The classical baseline beats the deep model, and the interesting part is the reason: with this little data the hybrid CNN-BiLSTM collapses toward the majority class.

01What it says

Deep learning needs data it didn’t have here. With 992 samples the CNN-BiLSTM learns to predict the majority class, while a TF-IDF Random Forest holds up better. A useful negative result, and a reminder to try the simple model first.

Co-developed with my classmate Arzoo Ghani Khan for the Machine Learning course.