Here’s a complete, well-structured README.md tailored for your project, assuming it's located in a GitHub repo and follows best practices:

---

# Knowledge Distillation for Neural Collaborative Filtering (NeuMF)

This project explores the effectiveness of **Knowledge Distillation** techniques in reducing the parameter size of the [NeuMF](https://arxiv.org/abs/1708.05031) model without significant loss in recommendation performance. It is based on the **MovieLens 100K** dataset and was completed as part of the university course **"Deep Learning and Its Applications"**.

## 🔍 Objectives

- Reproduce NeuMF results on a different dataset (MovieLens 100K).
- Investigate the impact of architecture choices (e.g. number of MLP layers, embedding size).
- Compare NeuMF to NMF in terms of HR@10, NDCG@10, and model size.
- Implement and evaluate 3 different **knowledge distillation** strategies:
  - Response-based distillation
  - Feature-based distillation
  - Relation-based distillation

## 📁 Repository Structure

. ├── ex_12.py # Main training/evaluation script for student distillation ├── Dataset.py # Dataset loader and preprocessor ├── evaluate.py # Evaluation metrics (HR@K, NDCG@K) ├── NeuMF.py # NeuMF model definition ├── MLP.py # Standalone MLP (student model) ├── Pretrain/ # Folder with pre-trained teacher weights (.npy) ├── logs*/ # Output logs from distillation experiments ├── script_12.sh # Shell script to automate 10-run experiments └── Data/ └── ml-100k/ # MovieLens 100K data files

## ⚙️ Installation

Create a conda environment (recommended):
```bash
conda create -n ncf python=3.6 tensorflow=1.14 numpy pandas
conda activate ncf

Install additional dependencies:

pip install scikit-learn matplotlib

📌 Notes

• All experiments are repeated 10 times, and results are reported as mean ± std.
• Negative sampling (--num_neg) and architecture settings are configurable.
• Pretraining is optional; this project focuses on training from scratch + distillation.

📚 References

• Xiangnan He et al. "Neural Collaborative Filtering", WWW 2017.
• Hinton et al. "Distilling the Knowledge in a Neural Network", NIPS 2015.
• Knowledge Distillation Benchmark: https://arxiv.org/abs/2006.05525

👤 Authors

• \Kleidonaris Dimitris – MSc Student, Department of Electrical & Computer Engineering, University of Thessaly
• \Aggelos Tzikas – MSc Student, Department of Electrical & Computer Engineering, University of Thessaly

📅 Submission