Recommendation System Model

Introduction

This repository contains a recommendation system model designed to provide personalized movie recommendations to users utilizing collaborative filtering techniques and the MovieLens dataset.

Purpose of the Project

Dive into the world of recommender systems, focusing on exploring different methods to use the data (such as negative sample generation and splitting), various evaluation techniques, and different model architectures. This project emphasizes clean code, modularization, and mastering Git/GitHub. Feel free to contribute and enhance the functionality and performance of the recommendation system 😊.

Model Functionality

Data Preprocessing

• Rating Thresholding: preprocesses the raw MovieLens data by defining a rating threshold to separate positive (relevant) user/item interactions from negative ones.
• Negative Sample Generation: balances the number of negative and positive samples by selecting dissimilar items (to the positive ones) based on metadata such as genre. This ensures a diverse set of negative samples for training.
• Dataset Splitting: divides the MovieLens dataset into training and testing subsets using either user-based or temporal splitting methods.

Hyperparameter Tuning

• Automated Search: performs automated hyperparameter tuning using the Hyperopt library, optimizing the model's performance based on predefined search spaces and evaluation metrics. This includes tuning the architecture of the collaborative filtering model (dot product, concatenated, deep learning, and hybrid models).

Evaluation

• Mean Average Precision at k (MAP@k): evaluates the model's performance using MAP@k, a metric that measures the average precision of the top-k recommendations provided to users.
• ROC Curve, Precision-Recall Curve, and Confusion Matrix: analyzes the ROC curve, precision-recall curve, and confusion matrix to assess the model's ability to discriminate between positive and negative items.
• Cosine Similarity: computes the cosine similarity between item embeddings to measure the similarity between movies. This metric helps assess how well the model captures the relationships between similar movies.
• Embedding Visualization: the model visualizes item embeddings using t-SNE (t-distributed stochastic neighbor embedding), which allows to explore the distribution and relationships between items in the embedding space.

Inference

• Loading the Model: loads the trained model and its weights from storage.
• Generating Predictions: predicts user-item interactions using the model.
• Filtering Recommendations: ensures no items already interacted with by the user are recommended.
• Providing Recommendations: outputs the top N recommendations for each user.