DESIGN AND DEVELOPMENT OF RECOMMENDATION SYSTEM USING GRAPH NEURAL NETWORK

Abstract

As current communication technology has advanced, e-commerce has become increasingly prevalent. These days, people can purchase goods from the comfort of their homes through online retailers like Amazon, Flipkart, and others, as well as find entertainment on websites like YouTube and Spotify. Another issue of "spoilt for choices" arises because internet platforms may accommodate a vast number of things from which users can choose. Recommender systems have become a very powerful tool to address this problem; given the user's past behaviour and the attributes of the various items, these systems apply an algorithm to create a candidate set from the entire set of items, rank the items in the candidate sets, and then display the items to the user based on the item ranking. We provide a thorough analysis of a recommender system in this dissertation. We go over the model that underpins it, its phases, the algorithms that drive these systems—whether they are sophisticated techniques based on deep learning and machine learning or more conventional approaches like matrix factorization—as well as the advantages and disadvantages of each. We also go over the applications of recommender systems. As the internet's reach increases and more people begin consuming digital material and engaging in e commerce, we have concentrated on session-based recommender systems. Both consumers and producers benefit from a session-based recommender system since it improves the buying experience for consumers and allows businesses to make better decisions by analysing the data produced by the system and using that information to improve the recommender systems' performance, which in turn improves the customer experience. We're putting forth a brand-new approach for session-based recommender systems. For our model, graphical neural networks (GNN) are being used. After preprocessing the user-item interaction datasets, our model learns item and positional embeddings, learns each item's relevant neighbourhood using item-KNN, generates a local and global graph, and obtains an embedding of each item in both local and global contexts. It then adds the local and global embeddings of the item to obtain the final representation, and uses the dot product of the final representation and the initial item embedding to obtain a final score that indicates the item's significance to the user. Lastly, it suggests items to the user depending on the final score. We have assessed our model's efficacy using precision (P@K) and mean reciprocal rank (MRR@K). Diginetica, TMall, and Nowplaying datasets were used. Our approach outperforms all non-graphical neural network-based recommendation systems. Compared to other GNN based models, ours performs significantly better.