This repository contains my graduate-level coursework and independent experiments focused on building, optimizing, and evaluating recommendation algorithms. The projects range from Deep Learning-based Collaborative Filtering using PyTorch to Content-Based systems leveraging NLP on massive unstructured datasets.
- Core: Python, PyTorch (GPU), Scikit-learn, Pandas, NumPy
- Techniques: Matrix Factorization, Neural Collaborative Filtering (NCF), TF-IDF, Cosine Similarity, Grid Search
- Optimization: AdamW, ReduceLROnPlateau, Early Stopping, Gradient Descent
- Engineering: GPU Acceleration, Pre-loaded GPU Tensors, Parquet Checkpointing
- Environment: Google Colab Pro (High-RAM & GPU Runtime)
Dataset: MovieLens 100K
This project benchmarks five distinct model architectures to minimize prediction error (RMSE) on sparse user-item interaction data.
- Architecture: Models were implemented in PyTorch. To maximize training speed, the entire dataset was converted to tensors and loaded directly onto the GPU to eliminate data transfer overhead.
- Optimization: Training utilized the AdamW optimizer (to decouple weight decay) and a ReduceLROnPlateau scheduler (halving the learning rate after 2 epochs of stagnation).
- Grid Search: An extensive search was performed across 125 combinations of embedding dimensions (8-128), learning rates, and regularization strengths.
- Hardware: Executed on Google Colab Pro to support intensive grid search iterations and GPU-resident training.
Models were evaluated using 5-Fold Cross-Validation. The optimal configuration was found to be Embedding Dim=32, LR=0.005, and Reg=1e-06.
| Model Architecture | Test RMSE |
|---|---|
| Matrix Factorization (w/ Bias) | 0.9602 |
| Matrix Factorization (No Bias) | 0.9640 |
| Neural Collaborative Filtering | 1.7150 |
| NCF (No Bias) | 1.6069 |
| Bias Only Baseline | 2.7879 |
Finding: The Matrix Factorization with Bias model achieved the lowest RMSE and fastest convergence, significantly outperforming the Neural Network (NCF) architectures on this specific dataset.
Figure 1: Validation RMSE over 100 epochs. The Matrix Factorization model (Green) converges faster than Neural Network architectures.
Figure 2: Impact of embedding size on error. Performance degrades at dimensions higher than 32, indicating overfitting.
Dataset: MovieLens Tag Genome 2021
The objective was to predict user ratings by analyzing textual content from over 2.6 million raw movie reviews.
- Data Pipeline: Raw reviews were aggregated by movie and cleaned (punctuation removal, stop words). The processed data was saved to a Parquet file checkpoint to bypass processing times on repeated runs.
- Memory Management: Due to the dataset size (52,000+ movies), standard similarity matrix calculations caused memory overflows. An "on-the-fly" calculation method was implemented to compute similarity scores only for necessary target pairs during the prediction loop.
- Hardware: Google Colab Pro High-RAM runtime was required to handle the large-scale text data processing.
- TF-IDF + Cosine Similarity: Weighs words by importance/frequency.
- Binary Counts + Jaccard Similarity: Weighs words by simple presence/overlap.
Models were evaluated on a sample of 5,000 ratings using Mean Absolute Error (MAE) and Hit Ratio.
| Model Strategy | MAE | RMSE | Hit Ratio (@10) |
|---|---|---|---|
| Binary + Jaccard | 0.7593 | 0.9850 | N/A |
| TF-IDF + Cosine | 0.7607 | 0.9855 | 0.20% |
| Global Average Baseline | 0.8355 | 1.0545 | - |
| Random Baseline | 1.5861 | 1.9565 | - |
Finding: The Binary + Jaccard model yielded the lowest error, suggesting that simple word presence was a more effective signal than term frequency for this specific review dataset.