TEAN: Timeliness enhanced attention network for session-based recommendation

Abstract

Session-based recommendation task attracts more researchers’ attention in recent years. However, previous approaches suffer from limited timeliness since they overlook dynamic features of items and temporal semantic information, which results in inappropriate prediction. In this study, we propose an attention-based model named Timeliness Enhanced Attention Network (TEAN). It first extracts features of user and item from static and dynamic perspectives and then employs temporal semantic information by a time-cross mechanism. Our model is capable of ranking items based on timeliness enhanced features. Besides, we apply a pre-training method based on word2vec to learn embedding vector for users, items and temporal semantic information in an elegant way. Experiments on three datasets of different domains demonstrate that our approach improves performance opposed to other methods.

Introduction

Session-based recommendation system [1](SBRS), as one of sequential recommendation tasks, has been proposed to improve personality by understanding users’ dynamic and evolving needs. By tracking an ongoing session between a user and system, SBRS aims to predict next action(e.g. click on an item) based on recent interactions in that session.

Traditional methods based on collaborative filtering (CF) [2], [3], [4] that rely on long-term preference might fail to catch exact needs of users since they ignore the context of a session. For capturing sequential patterns, early Marchov Chain(MC) [5], [6], [7] based approaches have been proposed, which assume next action is only based on previous action (or several actions). Another method, recurrent neural network (RNN) [8], which builds up the context of previous actions via a hidden state, has led to vital progress. Both methods could be used for SBRS, while they are limited in some situations. The major problem for MC-based methods is that the state space becomes unmanageable when including all possible sequences of user selections over all items. For RNN, the hidden state it learns probably overlooks some factors related to a user’s main purpose. Recently, The transformer [9] model and its main method named self-attention, which is effective for extracting the inner relations of sequences, have achieved state-of-art performance. Inspired by this method, some attention-based models [10], [11] for sequential recommendation have been proposed, that are not only capable of encoding users’ historical actions into a context vector, but also finding out the key factors to the user’s intent.

Timeliness is significant for SBRS since user’s preferences, item’s popularity or characteristics, and temporal semantic information are always changing, which requires timely recommendation algorithms to catch these changes in time. However, existing methods suffer from limited timeliness when making recommendations: (1) Most methods consider an item’s features, but few of them explicitly utilize its dynamic features and represent it as static features (e.g. attributes features), which makes it difficult to predict when items have similar static features. In this situation, the dynamic features of items can be represented through their historical users who interacted with them. Then the item with more users who have similar preferences with the current user will be recommended. (2) Temporal semantic information about when interaction happens, also influences timeliness which existing models ignore. For example, well-timed items (e.g. coat in winter) are more likely to be interacted by users. Besides, users intend to interact with individual items in some festivals or days. In other words, recommendation algorithms should take into account the influences of temporal semantic information on users and items.

In this study, we propose a Timeliness Enhanced Attention Network to solve the above problems. The design of our model follows one principle: the right user meets the right item at the right time. In TEAN, we first use a word2vec [12] based pre-training method to get embedding vectors for users, items, and temporal semantic information. Then, we represent the right user by fusing static features and dynamic features of current session, using attention mechanism. The right item is represented similarly, but its dynamic features are reflected in the historical users who interacted with it. Next, we apply a time-cross mechanism to describe the influence of temporal semantic information on the user and the item. Finally, our model predicts the probability that the user interacts with the item. The dynamic features of items and time-cross mechanism improve timeliness of our model, which leads to better recommendations.

The contributions of this article are summarized as follows:

• We propose a novel framework for session-based recommendations, which employs attention mechanisms to capture user’s behaviors and item’s characteristics while applying time-cross mechanism to learn temporal information.

• Our proposed model obtains item’s dynamic features from its historical users and temporal semantic features from whole interaction data, which are integrated into attention networks, resulting in enhanced timeliness.

• We carry out extensive experiments on three benchmark datasets. Several detailed comparative experiments are performed to demonstrate the merits and advantages of TEAN.