Recommender System Using Machine Learning And Deep Learning Techniques

With the advancements in modern communication technology, ecommerce has gained immense popularity. Today, there are online stores like amazon, flipkart etc. where people can buy stuff from the comfort of their home, entertainment venues like youtube, spotify etc. where they can get entertained and so on. Since, online platforms can host a large volume of items from which people can select, it presents another problem of “spoiled for choices”. To address this issue, recommender systems have emerged as a very potent tool. These systems take into consideration the user’s past behaviour as well as the attributes of the various items and after applying some algorithm it generates a candidate set from the complete set of items, it then ranks the items in the candidate sets and then present them to the user according to the ranking of the items. In this dissertation, we present a comprehensive overview of a recommender system. We discuss the model behind it; the phases it has; algorithms that power these systems whether they are traditional like matrix factorization or modern techniques based on ma chine learning and deep learning, their benefits and challenges associated with each; and the applications of recommender systems. We have focussed our attention on session-based recommender systems as the reach of internet is growing more-and-more people have started consuming digital content as well as pursuing ecommerce. A session-based recommender system is a win-win solution for both the consumers and the producers as the consumers get a better purchasing expe rience and businesses can optimize their decision-making with the analyses of the data generated by recommender system and using that same data to boost the performance of the recommender systems which in-turn enhances consumer experience. We are propos ing a novel method for session-based recommender system. We are using graphical neural networks (GNN) for our model. Our model takes the datasets of user-item interactions, preprocess them, then it learns item embeddings and positional embeddings, learn relevant neighbourhood of each item using item-KNN, creates local graph as well as global graph and gets an embedding of every item in each local as well as global context, then add the local and global embedding of the item to get final representation, then it takes a dot product of the final representation and initial item embedding to get a final score which signifies the importance of the item to the user. Finally, it recommends item based on the final score to the user. We have used precision (P@K) and mean reciprocal rank (MRR@K) for evaluating the effectiveness of our model. We have used Diginetica, TMall, Nowplaying datasets. Our method performs much better than any non-graphical neural network based recommneder system. Among the models based on GNN, our model performs much better.