Privacy-Preserving Sequential Recommendation: Enhancing User Privacy with Differentially Private RNNs, GRUs, and Transformers

A sequential recommender system predicts the users' next preferences or actions based on the sequential patterns in their past interactions. The project builds a session-based recommendation system with RNNs, GRUs, and Transformer models to make predictions about short-term user behaviors. It uses advanced preprocessing techniques and a Python-based software stack for model implementation and assessment. For ensuring user privacy, differential privacy mechanisms, including the noise addition during training and secure data aggregation are integrated, protecting sensitive information while maintaining the performance of model. The model are fine-tuned to capture complex patterns in user interactions while balancing accuracy, efficiency and privacy. Experimental results indicate that the best model attains an optimal recall of 0.1625 and mean reciprocal rank (MRR) of 0.1262 at ε = 5, which proves its efficiency in giving accurate recommendations while maintaining privacy.