Heterogeneous Information Crossing on Graphs for Session-based Recommender Systems

2025-05-06 0 0 3.34MB 25 页 10玖币

侵权投诉

Heterogeneous Information Crossing on Graphs for Session-based

Recommender Systems

XIAOLIN ZHENG, College of Computer Science, Zhejiang University, China

RUI WU, College of Computer Science, Zhejiang University, China

ZHONGXUAN HAN, College of Computer Science, Zhejiang University, China

CHAOCHAO CHEN*, College of Computer Science, Zhejiang University, China

LINXUN CHEN, MYbank, Ant Group, China

BING HAN, MYbank, Ant Group, China

Recommender systems are fundamental information ltering techniques to recommend content or items that meet users’ personalities

and potential needs. As a crucial solution to address the diculty of user identication and unavailability of historical information,

session-based recommender systems provide recommendation services that only rely on users’ behaviors in the current session.

However, most existing studies are not well-designed for modeling heterogeneous user behaviors and capturing the relationships

between them in practical scenarios. To ll this gap, in this paper, we propose a novel graph-based method, namely

eterogeneous

nformation

rossing on

raphs (HICG). HICG utilizes multiple types of user behaviors in the sessions to construct heterogeneous

graphs, and captures users’ current interests with their long-term preferences by eectively crossing the heterogeneous information on

the graphs. In addition, we also propose an enhanced version, named HICG-CL, which incorporates contrastive learning (CL) technique

to enhance item representation ability. By utilizing the item co-occurrence relationships across dierent sessions, HICG-CL improves

the recommendation performance of HICG. We conduct extensive experiments on three real-world recommendation datasets, and the

results verify that (i) HICG achieves the state-of-the-art performance by utilizing multiple types of behaviors on the heterogeneous

graph. (ii) HICG-CL further signicantly improves the recommendation performance of HICG by the proposed contrastive learning

module.

CCS Concepts: •Information systems →Recommender systems.

Additional Key Words and Phrases: Session-based recommendation, graph neural network, heterogeneous information

ACM Reference Format:

Xiaolin Zheng, Rui Wu, Zhongxuan Han, Chaochao Chen*, Linxun Chen, and Bing Han. 2018. Heterogeneous Information Crossing

on Graphs for Session-based Recommender Systems. J. ACM 37, 4, Article 111 (August 2018), 25 pages. https://doi.org/XXXXXXX.

XXXXXXX

*Corresponding author.

Authors’ addresses: Xiaolin Zheng, College of Computer Science, Zhejiang University, Hangzhou, China, xlzheng@zju.edu.cn; Rui Wu, College of

Computer Science, Zhejiang University, Hangzhou, China, CS_wurui@163.com; Zhongxuan Han, College of Computer Science, Zhejiang University,

Hangzhou, China, zxhan@zju.edu.cn; Chaochao Chen*, College of Computer Science, Zhejiang University, Hangzhou, China, zjuccc@zju.edu.cn; Linxun

Chen, MYbank, Ant Group, Hangzhou, China, linxun.clx@antgroup.com; Bing Han, MYbank, Ant Group, Hangzhou, China, hanbing.hanbing@alibaba-inc.

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not

made or distributed for prot or commercial advantage and that copies bear this notice and the full citation on the rst page. Copyrights for components

of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to

redistribute to lists, requires prior specic permission and/or a fee. Request permissions from permissions@acm.org.

Manuscript submitted to ACM

Manuscript submitted to ACM 1

arXiv:2210.12940v1 [cs.IR] 24 Oct 2022

2 Xiaolin Zheng, Rui Wu, Zhongxuan Han, Chaochao Chen*, Linxun Chen, and Bing Han

1 INTRODUCTION

view view view purchase view add to cart

timestamp

view

Fig. 1. An example session of user sequential behaviors in the e-commerce scenario.

Nowadays, recommender systems (RS) play a critical role in many real-life services such as e-commerce, streaming

platforms, and social networks [

]. Conventional recommender systems usually assume that long-term user proles

and user-item historical interactions are available [

]. However, in many practical scenarios, user identication

and historical information may not be available (except those in the current session) due to certain reasons, e.g., privacy

regulation [

]. Recently, session-based recommender systems (SBRS) are proposed to overcome this problem,

which utilizes the limited historical behaviors in each ongoing session to learn customer preferences and provide

recommendations [7,37].

Due to the highly practical value, researchers have proposed kinds of session-based recommendation methods in the

past few years. Most existing methods employ Markov chains [

] or recurrent neural networks (RNN) [

]

to infer user preferences from the transitional patterns in their interacted items. First, Markov chains based models

assume that a user’s subsequent behavior depends on the items he interacted in the latest behavior. However, it assumes

that independence only exists between subsequent behaviors and thence neglects the high-order relationships within

the behavior sequence, which limits the recommendation performance. Secondly, to relieve this assumption, researchers

have proposed RNNs-based models to capture the sequential properties, which leads to signicant progress. Among

them, GRU4Rec [

] is the rst work in session-based recommendation that applies gated recurrent unit (GRU) to

capture transitional patterns in user behavior sequence. By introducing an attention mechanism to capture the user’s

primary purpose in the current session, NARM [

] further improves the eects of GRU4Rec. STAMP [

] also achieves

good performance by applying an attention mechanism with simple multi-layer networks to capture users’ potential

preferences and current interests. Recently, researchers have utilized graph neural networks (GNN) to achieve state-

of-the-art performance across multiple elds [

]. There are also attempts, such as SR-GNN [

], GC-SAN [

], and

others [24,25,44,47], apply GNNs to improve the recommendation performance by modeling the complex transition

relationships between items.

Although the above mentioned methods have made signicant progress on SBRS, they focus on modeling a single type

of user interaction and fail to incorporate the multiple types. In reality, SRBS can collect various types of interactions,

such as browsing, adding-to-cart, purchasing, dwelling time, and ratings. In fact, various types of behaviors are essential

for SRBS to model a user’s preference in the current session, and ignoring this will inevitably result in estimation bias.

Figure 1 shows an example session in e-commerce scenario where the current user has three types of behaviors, including

Manuscript submitted to ACM

Heterogeneous Information Crossing on Graphs for Session-based Recommender Systems 3

view, add-to-cart, and purchase. Although this user may require earphones (or headphones), as he browsed extensively

at the beginning, the purchase of AirPods and his subsequent behaviors on the MacBook products indicate that his need

for earphones has been satised. In short, there are two things that should be noticed by SBRS: 1. The act of purchasing

represents the current user’s satisfaction with the need for headphones; 2. The subsequent behaviors of laptops indicate

the user’s new interests or demands, which should be taken care of by SBRS, and the importance of those behaviors

should be dierent from previous ones in the following recommendations. Ignoring multiple types of user behaviors

will fail to capture a user’s interest drift behind this signal and nally encounter poor performance when repeatedly

recommending the same type of items, e.g., earphones in this example. In recent years, a few SRBS work [

]

has been proposed to utilize multi-behavior in given sessions. However, those methods use heterogeneous information

by either combining individual behavior prediction tasks or simply concatenating the representations that are extracted

from dierent types of behaviors. Without considering the underlying relationships between heterogeneous behaviors,

these methods still have limited performance.

To adequately model the complex sequential relations between users’ interacted items and capture the relationships

among dierent types of behaviors, we propose a novel session-based recommendation method called

eterogeneous

nformation

rossing on

raphs (

HICG

). HICG integrates various types of user behaviors on heterogeneous graphs

and explores users’ current interests with their preferences on these graphs simultaneously. Specically, for each session,

HICG rst builds a heterogeneous item relationship graph, where the relationships between items are determined by the

types of behaviors. Then HICG learns the meanings of distinct behaviors by utilizing Gate-GNN [

] on the constructed

heterogeneous graph. After that, we propose a heterogeneous information crossing module in HICG to model a user’s

current interests and his long-term preference. More specically, this module models users’ current interests and

long-term preferences by utilizing an intra-behavior attention layer, which is used to capture the relationships between

their previous behaviors and most recent actions. In order to obtain a more precise representation of users’ current

interests, this module also applies an inter-behavior attention layer to characterize the relationships between dierent

behavior types. Finally, HICG combines current interests with long-term preferences to make recommendations. In

addition, inspired by the eectiveness of contrastive learning (CL) techniques in many areas [

], we propose an

enhanced version, named HICG-CL, which utilizes an extended CL module to improve the performance of HICG. This

module, in particular, uses the item co-occurrence relationships across dierent sessions to build a union graph, and

then applies a CL task on the graph to improve the learning of item representations. HICG-CL combines HICG with the

CL module in a unied architecture, and unies task and CL task under a multi-task learning framework.

The main contributions in this work are summarized as follows:

•

We propose a novel method HICG for session-based recommendation, which can utilize multiple types of user

behaviors to capture users’ current interests and their long-term preferences through heterogeneous graph modeling.

•

We develop a heterogeneous information crossing module, which can eectively learn the relationships among the

various behaviors with dierent user behavior types.

•

We also propose an enhanced version, named HICG-CL, which incorporates CL to enhance the learning of item

representations.

•

We conduct extensive experiments on three real-world datasets, and the state-of-the-art performance demonstrates

the superiority of our proposed methods.

Manuscript submitted to ACM

4 Xiaolin Zheng, Rui Wu, Zhongxuan Han, Chaochao Chen*, Linxun Chen, and Bing Han

The rest of this article is organized as follows. Section 2 reviews the related work on existing session-based recom-

mendation methods. Section 3 presents our proposed HICG and HICG-CL in details. Section 4 analyzes the results of

extensive experiments we conducted on three real-world datasets. Section 5 gives the conclusion of this article.

2 RELATED WORK

This section reviews the related work on SBRS, including SRBS with identical type behaviors modeling, SRBS with

heterogeneous behaviors modeling, and previous work which utilizes contrastive learning for SRBS.

2.1 Identical Type Behaviors Modeling for SRBS

Conventional methods are mainly based on collaborative ltering (CF) techniques and Markov chain (MC) sequential

models. CF based methods provide recommendation services based on similarity. For example, IKNN [

] rst compares

the items’ co-occurrences in dierent sessions to determine the session similarity, and then recommends objects in the

top-K similar sessions that the current user has not touched. MC based methods assume that user’s next behavior is only

related to the behaviors of his current moment. According to this assumption, researchers apply the rst-order Markov

chain to model user’s transfer between various interacted items, based on which session-based recommendation can

be made by simply computing the transition probabilities between items [

]. FPMC [

] combines CF and MC

approaches in a hybrid recommendation framework, and learns users’ long-term preferences with short-term interests

to predict the items in their next shopping baskets.

Neural network based models [

] have been proposed to model the sequential behaviors in SBRS with great success.

For example, GRU4Rec [

] rst employs recurrent network technology with Gated Recurrent Unit (GRU) to model

item interaction sequences. Then, GRU4Rec+ [

] further enhances the performance of GRU4Rec by introducing a data

augmentation technique. In addition, GRU4REC-DWell [

] improves GRU4Rec by applying the duration of each user

behavior in sessions. However, the recurrent neural network based models fail to learn the associations between users’

long-term behaviors and their current interests. To address this issue, researchers have investigated models based on

attention mechanisms, which can capture the relationship between users’ latest actions and their long-term behaviors.

By applying an attention mechanism, NARM [

] computes the importance scores of each historical behavior. After that,

the scores are used by NARM to calculate a weighted sum of behavior representations extracted by recurrent network,

and the calculation result is used to represent the user’s current intent. STAMP [

] emphasizes the importance of the

last click by using a global attention network. A trilinear product decoder is also used to improve its expressive ability,

which can describe the relationship between users’ long-term preferences, short-term interests, and the candidate items.

Inspired by Transformer [

], SASRec [

] learns the associations between dierent behaviors in sessions by applying

a multi-head self-attention mechanism, which also achieves great improvement. With consideration of collaborative

information in sessions and dynamic presentation of items, CoSAN [

] enhances the recommendation performance by

investigating neighborhood sessions.

Recently, researchers have introduced Graph Neural Network (GNN) technology to model the complex relationships

in graph-structured data, achieving excellent performance in a variety of tasks [

]. According to previous work on

SRBS [

], the association between items can be used to construct directed graphs, and using GNN on those

graphs can eectively generate session representations. SR-GNN [

] is a representation work among them. SR-GNN

generates a directed graph of interactions for each session, where the edges represent the order of adjacent interactions

within the given session. After that, SR-GNN processes each session graph to obtain the behavioral representations

with an attention mechanism. GC-SAN [

] improves SR-GNN by introducing a self-attention mechanism, which

Manuscript submitted to ACM

Heterogeneous Information Crossing on Graphs for Session-based Recommender Systems 5

can dierentiate the contributions of distinct adjacent nodes in the graph. By applying a global attention mechanism,

TAGNN [

] calculates the contributions of users’ previous interactions and uses them to enhance the representations of

their current interests. FGNN [

] and SGAT [

] utilise graph attention networks (GAT) to extract node associations with

multi-head attention. Both of them achieve a good breakthrough by eectively aggregating neighbors’ representations

as the target representation in the session.

Although the above mentioned methods have made remarkable progress on SBRS, these approaches only model

a single type of user interactions. In contrast, by building a heterogeneous item relationship graph and learning the

relationships among the user’s heterogeneous actions, our proposed HICG in this paper utilizes various user behaviors

to capture users’ current interests and long-term preferences.

2.2 Heterogeneous Behaviors Modeling for SRBS

To improve recommendation performance, SRBS methods with heterogeneous behavior modeling employ a variety

of user interaction types. With the consideration of distinct relationship types between items, those methods can

model users’ preferences more eectively. Although some early recommendation studies [

] have utilized the

heterogeneous information of user interactions to improve performance of traditional recommendation tasks, only a

few studies in SRBS have investigated the approaches to generating the representations from multi-behavior sessions.

Among them, MKM-SR [

] enhances the representations of the current session by incorporating user micro-behaviors

and it also involves KG embedding learning as an auxiliary task to promote the recommendation eects. MCPRN [

]

utilizes an RNN-based mixed channel purpose routing network to model the interacted items in the sequence, where

dierent channels are used to learn the distinguish interests. After that, MCPRN generates the recommendation results

by the integrated representation from all the channels. MGNN-SPred [

] applies auxiliary behaviors with target ones by

constructing a multi-relational item graph, which greatly improves the eectiveness of predicting the user’s subsequent

target behavior. Although the aforementioned methods have improved recommendation performance in SRBS, they use

heterogeneous behaviors without considering the underlying relationships between them. Unlike these methods, our

proposed HICG makes better use of heterogeneous behaviours and captures their inherent relationships.

2.3 Contrastive Learning for SRBS

Contrastive learning (CL) is one of the hot areas of recent scientic research, which belongs to the discriminative

self-supervised learning method. Compared with generative self-supervised learning (SSL) methods such as GAN [

]

and VAE [15], CL focuses more on learning the general characteristics of instances by distinguishing between similar

and non-similar entities [

]. To date, CL has a lot of studies in computer vision and natural language processing.

However, only a few studies investigated the ecacy of using this approach in SRBS. DHCN [

] models the session-

based data as a hypergraph and introduces CL as an auxiliary task to enhance the supergraph modeling ability, which

maximizes the mutual information between the session representations learned by two dierent networks. MHCN [

]

uses the aggregation of high-order user relations to enhance social recommendation. To alleviate the weakness of

feature dierences caused by the aggregation of dierent levels of information, MHCN utilizes CL to obtain the

connectivity that maximizes the hierarchical mutual information. COTREC [

] generates subgraphs by randomly

dropping edges in the session graph and treats homologous subgraphs as identical sessions in model learning. To obtain

a general representation of the given session, COTREC applies CL to dierentiate these subgraphs. Dierent from

the above methods that use CL to augment session representations, HICG applies CL to enhance the learning of item

representations by their co-occurrence relationships across dierent sessions.

Manuscript submitted to ACM

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

HeterogeneousInformationCrossingonGraphsforSession-basedRecommenderSystemsXIAOLINZHENG,CollegeofComputerScience,ZhejiangUniversity,ChinaRUIWU,CollegeofComputerScience,ZhejiangUniversity,ChinaZHONGXUANHAN,CollegeofComputerScience,ZhejiangUniversity,ChinaCHAOCHAOCHEN*,CollegeofComputerScience,Zhejiang...

展开>> 收起<<

Heterogeneous Information Crossing on Graphs for Session-based Recommender Systems.pdf

共25页,预览5页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Heterogeneous Information Crossing on Graphs for Session-based Recommender Systems

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: