ABSTRACT
Recommender systems are widely used to offer personalized recommendations for goods or services. Machine learning algorithms have been increasingly adopted by these systems in recent years, and selecting the appropriate algorithm is crucial. However, little guidance is available regarding the current usage of algorithms in recommender systems, and there are many challenges to creating effective recommender systems using machine learning algorithms. In this project, we build an advanced recommendation system using Neural Collaborative Filtering (NCF) and Variational Autoencoders (VAEs). We leverage a large-scale rating dataset containing user ID, movie ID, rating, and timestamp attributes to train and test our recommendation system. Our investigation focuses on combining traditional collaborative filtering and matrix factorization techniques with advanced deep learning-based techniques, such as NCF and VAEs, to build our recommendation system. We evaluate the performance of each model using multiple evaluation metrics, including precision, recall, and F1-score and compare their performance to decide on the best model based on recommendation quality, evaluation metrics, and time efficiency.
​
FUTURE WORK
Our study presents various opportunities for future work in movie recommendation systems. First, we could explore the combination of collaborative filtering and content-based recommendation systems to create a hybrid model that leverages the strengths of both approaches. This could potentially improve the accuracy of recommendations by incorporating additional features such as movie genre, director, and actors, which can be extracted from the movie metadata.
Incorporating contextual information: In this project, we only used user-item interaction data for the recommendation. However, incorporating contextual information such as time, location, and user demographics could improve the quality of recommendations.
Exploration of Deep Learning models: Although we compared the performance of two deep learning models in this project, there are many other deep learning models that could be explored, such as Convolutional Neural Networks (CNNs) and Transformer-based models.
Explainability and transparency: Recommendation systems can sometimes be seen as a "black box" where users do not know how the system arrives at its recommendations. Future work could focus on developing more transparent and explainable models to increase user trust and confidence in the recommendations.
Active Learning: The project evaluated different recommendation algorithms based on the accuracy of their predictions. However, accuracy alone may not always be sufficient for some applications. Future work could investigate active learning techniques to optimize the trade-off between recommendation accuracy and diversity.
Privacy and Security: With increasing concerns about data privacy and security, future work could focus on developing more privacy-preserving and secure recommendation systems, such as Federated Learning, Differential Privacy, and Secure Multi-Party Computation.
Multi-Objective Optimization: In this project, we mainly focused on optimizing for the accuracy of recommendations. However, there are often multiple objectives that need to be considered, such as diversity, novelty, and serendipity. Future work could explore multi-objective optimization techniques to balance these competing objectives with the system.
