Item Prediction with RNN Using Different Types of User-Item Interactions

Abstract

This paper deals with the session-based recommendations of different types of user-item interactions. Every user session includes sequences of item interactions such as item viewing, putting into basket and purchasing. Sequences that are constituted short events make the recommendation problems more challenging. Therefore, we applied a powerful state of the art Recurrent Neural Networks (RNN) with Gated Recurrent Unit (GRU) to train data and to predict the next item to be purchased. In the proposed method, Node2Vec representations of items are obtained using the probabilities of different useritem interactions. These representations are used as the initial weights of the GRU inputs. Experimental results are obtained on nearly one million sessions that are constituted by view, basket and purchase interactions which were collected from a Turkish e-commerce website. Experiments that are evaluated by using Mean Reciprocal Rank (MRR) and Recall metrics show that the proposed method performs 63% Recall and 41% MRR results.