END TO END GRAPH NEURAL NETWORKS BASED RECOMMENDER SYSTEM

Abstract

Two extensions based on Graph Encoder-Predictor Networks are presented in the thesis named ”GNN-based Recommender systems” to tackle the difficult task of Recommender systems. Recently, recommender systems capable of interpreting graph-like data struc tures to generate meaningful representations from user interactions and social network data have been designed by using neural networks. The initial expansion suggested in this thesis uses Graph-Norm, an alternative normaliza tion method to the ones already in use. The second extension replaces the conventional methods such as concatenation with Mem Pooling Aggregation. Both approaches use the Encoder architecture’s power to create suggestions that take advantage of social interac tions. By suggesting a recommendation system that may utilize the underlying social ties be tween users and objects, this research effort is intended to remedy this issue.Metrics like MAE and RMSE for regression-based systems are effectively used to assess the net effect of the presented methodologies. The results show how well the strategies produce recom mendations that are deemed adequate given the limitations. The thesis proves the viability of GNN-based recommendation algorithms in the end. The outcomes confirm the possibility of the suggested techniques for producing reliable, high-quality recommendations from user-item associations in the data. The results un derline the importance of including Graph-Norm and Mem Pooling aggregation in the Graph neural network design. Future research can investigate various loss functions, net work designs, and larger datasets for further validation. The proposed approach has the potential to be used in a variety of real-world situations.