Current Interest Enhanced Graph Neural Networks for Session-based Recommendation

Abstract

Session-based recommendation, aimed at predicting the unknown part of the anonymous interactive session, is formed via the known interaction information between users and items. Existing studies are limited to mine the session characteristics between items to detect the interest of users for the moment through treating session sequences as graphs. Additionally, as the interactive time sequence of items can reveal the recent changes in the user's interest behavior, the seriality of the session will lose after being modeled as a graph structure. This paper proposes a Current Interest Enhanced Graph Neural Network(CIE-GNN) model for the session-based recommendation. By design, the session sequences are represented by directed graphs, in an attempt to harness the graph neural network's power to capture rich item transitions in sessions. Then we take the last click in the original sequence as the query employed in the multi-head mechanism to adaptively acquire the user's current interest expression more accurately. At the same time, position embedding is used to combine conversation's sequential nature, to leverage both sequential perceptron and graphic perceptron models. Numerous researches concerning two genuine datasets exhibit that this method is predominant to other state-of-the-art methods reliably.