APWeb _Noise-Resistant Graph Neural Networks(1).pdf

墨天轮福利君

17页

0次

2024-10-29

免费下载

Noise-Resistant Graph Neural Networks

for Session-Based Recommendation

Qi Wang

,AnbiaoWu

,YeYuan

)

,YishuWang

Guangqing Zhong

, Xuefeng Gao

, and Chenghu Yang

Beijing Institute of Technology, Zhuhai, China

School of Computer Science and Engineering, Northeastern University,

Shenyang, China

wangyishu@mail.neu.edu.cn

School of Computer Science and Technology, Beijing Institute of Technology,

Beijing, China

yuan-ye@bit.edu.cn

Fabarta Co., Ltd., Beijing, China

{jameszhong,andy,ych}@fabarta.com

Abstract. Session-based recommendation have received increasing

attention due to the importance of privacy and user data protection, aim-

ing to predict the next click of a user based on a short anonymous inter-

action sequence. Previous works have focused on users’ long-term and

short-term preferences, ignoring the noise problem in session sequences.

However, session data is inevitably noisy, as it may contain incorrect

clicks that are inconsistent with the user’s true intent due to mislead-

ing product information. Therefore, in this paper, we propose a novel

framework called N

oise-Resistant Graph Neural Networks (NRGNN) to

address the noise problem in session-based recommendation. NRGNN

innovatively introduces two key components: Noise-Resistant Graph

Contrastive Learning (NR-GCL) and Cross-Session Enhanced Short

Preference (CS-SP). NR-GCL is a graph contrastive learning method

that employs minor perturbation augmentation to reduce the impact of

noise problem in the entire session on the accuracy of the results. CS-

SP utilizes cross-session information, aiming to address the problem of

poor recommendation accuracy when the last item is noisy. To evaluate

our proposed method, we conduct comprehensive experiments on three

real-world datasets. The experimental results demonstrate that NRGNN

outperforms the state-of-the-art methods.

Keywords: Session-based Recommendation

· Graph Neural

Networks

· Graph Contrastive Learning

1 Introduction

Session-based recommendation involves using the current click sequence of an

anonymous user to make predictions about his/her next click when historical

 The Author(s), under exclusive license to Springer Nature S ingapore Pte Ltd. 2024

W. Zhang et al. (Eds.): APWeb-WAIM 2024, LNCS 14962, pp. 144–160, 2024.

https://doi.org/10.1007/978-981-97-7235-3

_10

Noise-Resistant Graph Neural Networks for Session-Based Recommendation 145

data and identifying information about the user are not available, especially for

users who are not logged in or newly registered [4,6,20]. In addition, session-

based recommendation is capable of capturing dynamic preferences that may

diﬀer from the user’s historical preferences [17]. As the session recommenda-

tion does not require any information other than clicks, it can well protect user

privacy and data security and has wide application in e-commerce, music and

streaming media [4,6]. The ﬂow of session-based recommendation is depicted in

Fig. 1, where the user’s interest in music-related products is evident and the next

likely click is a headphone. Therefore, the aim of session-based recommendation

in this scenario would be to recommend a suitable music headphone for the user.

Fig. 1. The ﬂow of session-based recommendation

Many existing work treat user preferences as a combination of long-term and

short-term preferences. Long-term preferences are learned from the entire session

and may represent the user’s broader interests, such as the music-related prefer-

ences seen in Fig. 1. Short-term preferences, on the other hand, are learned from

the last item in the session and may indicate the user’s immediate intent, such

as the expected click on headphones in Fig. 1. Existing state-of-the-art methods

typically utilize Graph Neural Networks (GNNs) to model session data in order

to capture users’ long-term preferences [2,18,21,22,24]. As for short-term prefer-

ences, most of the works consider the embedding of the last item in the session as

the short-term preferences, following the Markov Chain principle [10,16]. How-

ever, these approaches primarily focus on users’ long-term and short-term prefer-

ences in the session-based recommendation, overlooking an crucial issue, i.e., the

noise in session-based recommendation. The noise is often caused by misclicks

from users who may be misled by item information such as images or titles. An

illustrative session sequence s

in Fig. 1, presents an instance of a sequence with

noise, where the third term is a source of noise. On e-commerce platforms, gam-

ing headsets and Bluetooth headsets often exhibit visual similarities, leading to

user misclicks and the generation of session sequences contaminated by noise.

Models that lack robustness or noise immunity exhibit a signiﬁcant decrease in

performance when dealing with such noisy session sequences.

In order to reduce the impact of noise on the accuracy of session-based rec-

ommendation as much as possible, we propose a novel framework called noise-

resistant graph neural networks (NRGNN) for session-based recommendation.

Inspired by the fact that the graph contrastive learning (GCL) can be eﬀective

146 Q. Wang et al.

in enhancing the robustness of graph representation learning [23,30], NRGNN

uses GCL to reduce the eﬀect of noise in sessions while improving the perfor-

mance of graph representation learning. The common augmentation methods in

GCL involve randomly adding and dropping nodes and edges to achieve graph

augmentation. However, such augmentation method could disrupt the original

topology of the session graph in session-based recommendation. To address this

issue, we introduce a novel component in NRGNN, called Noise Resistant Graph

Control Learning (NR-GCL), which utilises a restricted random perturbation-

based augmentation method. Speciﬁcally, the augmentation method is imple-

mented by adding size-constrained perturbations to the embedding space of the

items in the session to perform graph augmentation.

Consider that traditional Markov Chain struggle to capture the accurate

short-term preferences when the noise occurs exactly at the last click. Therefore,

we propose another component in NRGNN called Cross-Session Enhanced Short

Preference (CS-SP). CS-SP is speciﬁcally designed to mitigate the impact of

noise when it appears at the last item of a session. We observe that many users,

despite initially clicking on an item inﬂuenced by misleading item information,

subsequently click on an item that aligns with their true intent. Across multiple

sessions, we typically ﬁnd adjacent clicks of items that reﬂect the true intent

alongside incorrectly clicked items. CS-SP dynamically aggregate the neighbors

clicks with user’s true intent across all sessions. This aggregation is accomplished

through a attention mechanism that takes into account the overall intent of

the session. By employing this adaptive aggregation strategy, CS-SP eﬃciently

captures the correct short-term preferences of users, outperforming traditional

Markov chain-based approaches.

To sum up, the main contributions of this paper are as follows.

– We propose a novel framework, called NRGNN, which eﬀectively addresses

the noise problem in session-based recommendation by NR-GCL and CS-SP.

To the best of our knowledge, NRGNN is the ﬁrst GNN-based method that

speciﬁcally tackles the problem of noise in session-based recommendation.

– We propose NR-GCL, which is based on restricted random perturbation, to

macroscopically address the noise problem of noise in the entire session and

improve recommendation accuracy.

– To address the problem that the last item is noisy in session, we propose CS-

SP, which further enhances recommendation accuracy by integrating infor-

mation from multiple sessions in session-based recommendation.

– We conduct comprehensive experiments on three real-world datasets and com-

pare NRGNN with existing state-of-the-art methods. The results demonstrate

that NRGNN achieves the highest recommended accuracy and exhibits a high

level of robustness to noise. The code of this paper is released at https://

github.com/QiWang98/NRGNN.

of 17

免费下载

arcgraph paper

相关文档

评论