CaSe4SR: Using category sequence graph to augment session-based recommendation

Abstract

Session-based recommendation aims to predict next item based on users’ anonymous behavior sequence within a short time. Recent studies focus on modeling sequential dependencies or complex relations among items in a session via recurrent/convolutional/graph neural networks. However, the following problems still remain: for short sessions, limited interactions cannot manifest user’s intent clearly; for long sessions, user’s interest may drift but be blurred by complex transitions. Motivated by the observation that different items are often belong to only a few categories or that closely related, in this article, we tackle these challenges by leveraging item category information, which is a concise form of knowledge and readily available in many platforms. We propose a novel method CaSe4SR that utilizes category sequence graph to augment session-based recommendation. In CaSe4SR, we build an item graph and a category graph, from user behavior sequence and item category sequence. The latter summarizes the former at concept level, which reduces item-level user behavior noises and makes user’s interest clearer. Afterwards, graph neural networks are applied on item graph and category graph respectively to learn representations of items and categories. Then two alternative fusion strategies and attention mechanism are designed to integrate them, yielding global embedding of the session, which is further combined with representation of last item to get ultimate session representation. Extensive experiments on real-world datasets show that CaSe4SR outperforms other state-of-the-art methods consistently. Detailed analysis reveals that category sequence graph is beneficial for next-item recommendation in sessions with different lengths.

Introduction

In the era of information overload, recommendation systems have played an increasingly important role in many online services and have become an important tool for people to filter information. Currently, recommendation systems have been applied in many scenarios. For example, we need a recommendation system to help us recommend products that are suitable for us when shopping online, and when listening to music, we need to check the recommendation list to find the songs we like. Most existing recommendation systems analyze the user’s historical clicks, purchases, ratings, and other interactive behavioral data and combine the user’s identity information to discover his/her preferences. According to users’ preferences, merchants can help users choose items which they are interested in from massive resources, while promoting the sales of items. However, in many application scenarios, for the purpose of protecting user privacy, we cannot obtain user identity information completely. And the user’s past activities cannot be recorded continuously, which results in that we can only obtain user behavior records within a certain time interval, as known as a session. The user’s session sequence can be a sequence of behaviors occurring in an hour, a day or even a few days [1]. For instance, those songs listened by a user in an hour or a day can be called a session. Session-based recommendation makes predictions of the next behavior only based on the latest behavior sequence, not on long-term user behaviors. And it does not require much user information, so it is more practical in this age when user information is protected, and has become a research hotspot recently.

The method originally used for session-based recommendation is to evaluate the similarities among items [2] based on their co-occurrences in the same session. However, in order to improve the efficiency of the algorithm, these methods always consider the last interactive item in a session and ignore the sequential information of the ongoing session, resulting in unsatisfactory recommendations. Then, Markov Chain-based methods [3], [4] appeared to solve the problem of session-based recommendation. They exploit sequential behavior data and predict user’s next behavior based on last behavior. But these methods only model the local sequential behavior and the required state space becomes unmanageable quickly when trying to calculate the probability on all items for all users. Recently, there are many algorithms based on deep learning for session-based recommendation tasks. Recurrent neural network has been proved very effective in sequence modeling. It can consider the information of the entire session sequence. Hidasi et al. [5] were the first to introduce recurrent neural networks into session-based recommendation task. Tan et al. [6] improved the performance of this model by using data augmentation techniques, while considering the characteristics of user behavior changes over time for recommendation. Li et al. [7] combined the attention mechanism with RNN, while capturing the sequential behavior characteristics and user’s main purpose. Although RNN-based methods can capture the sequential dependencies among behaviors in a session well, the transition mode among items in a session is usually not single-way, but more complicated. RNN-based methods cannot model the complex transition modes among behaviors. Recent years have witnessed significant developments in Graph Neural Networks(GNNs) [8], [9], which have proven to be helpful in making up for this shortcoming. For example, Wu et al. [10] used graph neural networks to model session sequences to capture complex transition relationships between behaviors, which achieved better recommendation performance than RNN-based methods.

Although the above methods all show good recommendation performance in some session-based recommendation tasks, there are still some challenges in this kind of problem. This is because there are usually limited behaviors in an anonymous session, and the above methods mainly focus on item-level modeling, which may lead to deviations in the perception of user interests. First of all, for short sessions, limited interactive behaviors cannot show user’s real intention. For instance, if a user clicks iPhone XR after vivo X30, from the perspective of the item alone, since they are two completely different items, it is difficult for us to get the user’s real intention from this session with only two behaviors. Secondly, because user’s interests are changeable, the transition relationship between behaviors is more complicated in long sessions and the complex transition mode will obscure our judgment on user’s preference. Therefore, we can model the session sequence by adding some auxiliary information to solve the above two problems. Perhaps this information can be obtained from other domains [11], [12]. But as we mentioned earlier, user identity information is not available in many cases and only user’s interactive records in a short time is available, which makes these methods unsuitable. So we can add information to the behavior sequence from the item side (such as item category) to assist us to capture user’s preference.

In our daily life, in order to facilitate users’ search, e-commerce platform such as Tmall usually assign a category label to each item. In this way, users can find the products they want to buy more quickly and accurately. From the perspective of the platform, the recommendation system will recommend similar products for users based on the products they have visited. For example, iPhone 11, iPhone XR and iPhone XS are all iPhones, just different models. But considering only from the item level, they have nothing in common and have different item IDs. If these three items constitute a session sequence $[iPhone11,iPhoneXR,iPhoneXS]$, it is easily affected by detail noise when it is modeled only at the item level. However, if we combine the analysis at the item level and the category level at the same time, we will be pleasantly surprised to observe that the user’s real intention is to find a mobile phone that suits him/her. In this way, we can accurately predict the items that meet the needs of users.

In this article, we take the category information of the item as supplementary information for the modeling of the session sequence, and propose a model named CaSe4SR, short for using Category Sequence graph to augment Session-based Recommendation. It can alleviate the cold start problem to a certain extent and help us get a more accurate expression of the session sequence. This makes the user’s behavior intention clearer, which in turn can provide the user with more satisfactory recommendation results. Specifically, we construct the category sequence corresponding to the item sequence as a directed graph, and use the graph neural networks to learn the representation of the category. Finally, we fuse the representations of category and item to jointly characterize the session. The main contributions of this work are summarized as follows:

(1) We introduce category information into session-based recommendation models and establish a category sequence model. We describe the behavior sequence at the conceptual level, thereby avoiding detailed noise. As a supplement to the behavior sequence information, the user’s intention will become clearer after adding category information.

(2) We construct a category session graph neural network model and learn the representation of the category, thereby capturing the non-sequential dependencies in the session.

(3) Based on the fusion of item and category knowledge representations, we use attention mechanism to weight the fusion representations to obtain the global embedding of the session.

(4) We conduct extensive experiments on real-world datasets. The results show that our model evidently outperforms the state-of-art methods.

The rest of the paper is as follows. After introducing the related works of session-based recommendation in Section 2, we show our proposed model in Section 3 and describe the model in detail. In Section 4, we give the experimental results and discussions. Finally, we conclude this paper and put forward the outlook for future work.