Week 3 – Deep Learning Model Development & Refinement

27 Jun 2025 - Hirra Asif

This week I have been testing different class balancing techniques ⚖️, including SMOTE, random oversampling and class weight adjustments to see if they can improve the performance of the machine learning models on social media data. I ran each method through 5-fold cross‑validation on the models to identify which approach works best.

I have also been writing the code to train the deep learning models, including a neural network and transformer models such as uncased BERTBASE and RoBERTa 🤖, and have implemented a training pipeline to record training progress and key metrics (e.g. loss, accuracy, F1‑score). 📊

➡️ Next, I will evaluate their performance on the test data, fine‑tune hyperparameters to compare them against baseline models, and begin looking at implementing interpretability techniques (for example SHAP and LIME).