Recommendation System by Link Prediction Approach on Transactional Data

Official URL: https://risc01.sabanciuniv.edu/record=b2579423_(Table of contents)

With the continuous digitalization of the world, massive amounts of data are produced every second. Therefore, recommending relevant items to users has become a more important task in many systems. For this purpose, transaction data sets can be exploited in recommendation systems to understand underlying user interests. Commonly used recommendation systems adopt collaborative filtering based approaches that utilize a user-item matrix based on users’ past activity. However, these methods may suffer from sparsity and scalability issues. In this thesis, a link prediction based recommendation system combining graph representation learning algorithms and gradient boosting classifiers for transaction data sets is proposed as a scalable solution. Proposed system generates a network where nodes correspond to users and items, and links represent the interactions between them. A use case scenario is examined on a credit card transaction data set as a merchant prediction task which is predicting the merchants where users can make purchases in the next month, in a link prediction context. Performances of common network embedding extraction techniques and classifier models are evaluated by conducted experiments, and based on these evaluations; the proposed system is constituted. A matrix factorization based alternative scalable recommendation method is compared with the proposed model. Proposed method has shown a superior performance than alternative method in terms of receiver operating characteristic curves, area under the curve, and mean average precision metrics.