A cost-sensitive temporal-spatial three-way recommendation with multi-granularity decision

Highlights

• We consider the temporality of recommendation process and spatiality of recommendation information for sequential recommendation.

• A RNN-based granulation method is designed to construct multilevel recommendation information in the space dimension.

• A temporal-spatial three-way recommendation (TS3WR) is presented to realize the multi-step recommendation in the time dimension.

• A RNN-TS3WR method is proposed to realize the temporal-spatial recommendation.

Abstract

In considering of the dynamic variations of user’s preference and item’s popularity, sequential recommender system (RS) has attracted much attention in recent years. In general, the sequential interactions between users and items will lead to both multilevel recommendation information (RI) in the space dimension and multi-step recommendation in the time dimension. To better capture the dynamic variations of user’s preference and reduce the recommendation cost, this paper proposes a novel sequential recommendation strategy from the temporal-spatial perspective. Firstly, in view of the temporality of user-item interactions, we design a granulation method based on recurrent neural network (RNN) to construct the multilevel RI. Then, in the light of the temporality of user’s preference and item’s popularity, we present a temporal-spatial three-way recommendation strategy (TS3WR) to realize the multi-step recommendation. Finally, by integrating the time factor with space factor, a temporal-spatial three-way recommendation based on recurrent neural network (RNN-TS3WR) is proposed to realize the recommendation with lower decision cost. Extensive experiments on three Movielens datasets verify the feasibility and effectiveness of our proposed methods, and demonstrate the advantage of our recommendation strategy in both recommendation cost and recommendation quality.

Introduction

In recent years, sequential recommender system (RS) has attracted much attention in recommender systems community due to the dynamic variations of both user’s preference and item’s popularity. Concretely, a user’s preference and taste will vary as the time goes on. For example, many young people who used to be the fans of iphone but now have switched to become the fans of Huawei or Samsung. In terms of this phenomenon, many sequential recommendation methods have been proposed to model the sequential behaviors of users. For instance, Rendle et al. [1] applied the Markov chain (MC) to model the sequential behaviors of users. Hidasi et al. [2] proposed a sequential approach based on gate recurrent unit (GRU) for session-based recommendations, where the basic GRU is modified to fit the task better by introducing session-parallel mini-batches, mini-batch based output sampling and ranking loss function. Tang and Wang [3] introduced convolutional neural network (CNN) into sequential recommendation models, and proposed a convolutional sequential embedding recommendation model to capture both general preferences and sequential patterns. Wu et al. [4] designed a session-based recommendation through graph neural network (GNN) to capture complex transitions of items through the session graph. Wang et al. [5] further proposed a novel session-based recommendation, namely, global context enhanced graph neural networks (GCE-GNN) to exploit the item-transitions over all sessions in a more subtle manner for better inferring the user preference of the current session. Compared with the conventional recommendation methods, i.e., content-based and collaborative filtering, such sequential methods are of great significance for precisely profiling a user or an item for more accurate recommendations [6].

However, existing sequential recommendation methods primarily concentrate on two-way recommendation under static environment. They ignore the cost-sensitive problem during the recommendation process, and seldom consider the multilevel RI and multi-step recommendation induced by the time. Observed by the limitations of two-way recommendation, three-way recommendation is proposed to deal with these problems. As an important application of three-way decision (3WD), three-way recommendation adopts the third recommendation option, deferment-recommendation, into RS and takes the cost-sensitive problem of recommendation error into account. Zhang et al. [7] employed three-way decision to RS and proposed three-way recommendation to minimize the average recommendation cost. Besides, Huang et al. [8] proposed a novel cost-sensitive three-way recommendation method by learning pair-wise preference. Nevertheless, previous studies on three-way recommendation primarily concentrate on the single-step recommendation under single recommendation information (RI), but ignore the temporality of RS. In general, a dynamic decision information system with temporality may naturally form a granular structure in the space dimension and dynamic decision process in the time dimension [9]. For instance, in online shopping platform, for a new user, system can not capture the preference of the user, thus the item with higher popularity will be recommended to the user at the beginning. After a period of time, according to the browsing records and purchase records of the user, system may learn that the user primarily concentrates on the women’s clothes. Consequently, the women’s clothes will be recommended the user frequently. Furthermore, with the increasing of history records, the system can further capture the clothing style that the user like, and recommend the clothes with some specific styles to the user. Obviously, as the increasing of history records, system can capture more fine preference of the user. From the viewpoint of granular computing, as the time goes on, the historical records will gradually increase and the RI learned from history records will gradually refine. The reason is the system can capture more valuable information as the increasing of history records. It means that the information contained in the RI will become more and more sufficient and accurate. Thus, the sequential records with temporal order can be processed to the multilevel RI with spatiality. In addition, as mentioned above, user’s preference and taste will change over time [6]. For example, in the summer, the user may want to buy a skirt; in the winter, the user may want to buy a down coat. Thus, to match the dynamic preference of user, system should make different recommendation decision at different time, which naturally lead to the multi-step recommendation in the temporality. For this situation, the temporality and the spatiality should be integrated in a unified framework to improve efficiency. Therefore, with consideration of both the temporality of recommendation process and the spatiality of recommendation information, the multi-step recommendation under multilevel RI seems more reasonable.

Intuitively, to realize the multi-step recommendation from the temporal-spatial perspective, we should firstly construct a multilevel structure of RI. Granular computing (GrC) [10], [11], as an effective structured approach to thinking, problem solving, and information processing, has been widely used in multilevel information processing. However, existing granulation methods, such as, fuzzy set [12], rough set [13], quotient space [14], feature selection [15] and feature extraction [16], [17], etc, only take the spatiality of information into account but seldom consider the temporality of original data. Therefore, to model the sequential interactions between users and items, it is critical to design a temporal-spatial granulation method to deal with the data with temporality. Fortunately, Recurrent neural network (RNN) models are powerful tools for modeling sequences, and they can extend flexibly and can incorporate various kinds of data [18]. Recently, on account of the prominent performance for sequential modeling, RNN models, such as, long short-term memory [19], [20], [21] and gate recurrent unit [22], have been widely used for sequential recommender systems [18], [23]. In general, RNN can process a sequence by recursively applying a transition function to it’s internal hidden state for each unit of the input sequence [24]. It means that RNN methods can not only process the sequential data with temporality, but also generate a sequence with spatiality simultaneously. Thus, it is desirable to introduce RNN model to build the multilevel RI from the temporal-spatial perspective. What is worth mentioning is, there are also other models can be used to hand sequential task, such as, Markov chain [1], convolutional neural networks [3], etc. Due to the prominent performance of RNN in sequential data, this paper primarily discusses RNN model.

In the context of multilevel RI, we should further design an effective temporal-spatial recommendation strategy to realize the multi-step recommendation. As one of the typical multi-step models of three-way decision (3WD) [25], sequential three-way decision (S3WD) [26] provides a useful strategy to combine the temporal and spatial framework together. Meanwhile, as the extension of 3WD, S3WD can solve the cost-sensitive problem effectively. Recently, S3WD has been widely applied to deal with the multi-granularity decision problems. Li et al. [16], [17] introduced the S3WD to address the cost-sensitive problem in face recognition. Yang et al. [27] combined two-way decisions with three-way decision by the Bayesian decision procedure and presented a unified framework of the sequential approach and multi-class three-way decision. Inspired by the notion of GrC and S3WD, Liu and Ye [28], [29] built a novel dynamic three-way recommendation model to address the limitations of static two-way recommendation. Zhang et al. [30] developed a more generalized model of S3WD to balance the misclassification cost and the time cost in classifications and decisions. Meanwhile, in the aspect of theories, Yang et al. [9], [31] used the notion of S3WD to interpret and integrate the temporality and spatiality of three-way granular computing, and further constructed a novel framework of sequential three-way granular computing based on temporal-spatial multi-granularity learning. Intuitively, S3WD can not only effectively solve the problem of cost-sensitive but also realize the multi-step decision in the multilevel environment. Inspired by the previous studies, we present a more generalized framework of temporal-spatial three-way decision (TS3WD), and introduce the notion of TS3WD into RS to realize the temporal-spatial multi-step recommendation.

In summary, to better capture the dynamic variations of user’s preference and reduce the cost caused by recommendation error, our work takes both of the temporality of recommendation process and the spatiality of recommendation information into account. With that, we propose a dynamic recommendation strategy under multi-granularity environment to overcome the limitations of conventional sequential recommendation. First, from the spatial perspective, a RNN-based granulation method is designed to build the multilevel information. Then, from the temporal perspective, a temporal-spatial three-way recommendation (TS3WR) strategy is presented to realize the multi-step recommendation. Finally, with the integration of the temporality and spatiality, a temporal-spatial three-way recommendation based on recurrent neural network (RNN-TS3WR) is proposed to realize the multi-step recommendation under multilevel RI. A series of experiments on three Movielens datasets verify the feasibility and advantage of our proposed granulation method, as well as show that our proposed TS3WR strategy can effectively reduce the recommendation cost and improve the recommendation quality.

The main contributions of this paper include:

• (1)We take the temporality of recommendation process and the spatiality of recommendation information into account simultaneously, which can better capture the dynamic variations of user’s preference and item’ popularity.

• (2)We design a temporal-spatial granulation method based on RNN. This granulation method takes the temporality of original data into account instead of only concentrating on the spatiality of information.

• (3)We present a more generalized model of TS3WD in terms of S3WD. By introducing TS3WD into RS, a TS3WR strategy is conducted to realize the multi-step recommendation in the time dimension, it not only extends the theory of 3WD from a more generalized view, but also enriches the applications of 3WD.

• (4)We propose a novel temporal-spatial recommendation method, namely, RNN-TS3WR, by integrating the RNN-based granulation method with TS3WR. This method can reduce the recommendation cost and improve the recommendation quality simultaneously.

The rest of the paper is organized as follows: Section 2 introduces some preliminary knowledge. Section 3 defines the task and illustrates the adaptation of long short-term memory. In Section 4, we propose RNN-based granulation method, recommendation strategy and RNN-TS3WR. Section 5 shows the experimental results and analyses on three Movielens datasets. Finally, the conclusion and further works are presented in Section 6.