Self-supervised Graph-based Point-of-interest Recommendation

2025-05-03 0 0 1.78MB 19 页 10玖币

侵权投诉

YANG LI, The University of Queensland, Australia

TONG CHEN, The University of Queensland, Australia

PENG-FEI ZHANG, The University of Queensland, Australia

ZI HUANG, The University of Queensland, Australia

HONGZHI YIN, The University of Queensland, Australia

The exponential growth of Location-based Social Networks (LBSNs) has greatly stimulated the demand for precise location-based

recommendation services. Next Point-of-Interest (POI) recommendation, which aims to provide personalised POI suggestions for users

based on their visiting histories, has become the prominent component in location-based e-commerce. Recent POI recommenders

mainly employ self-attention mechanism or graph neural networks to model the complex high-order POI-wise interactions. However,

most of them are merely trained on the historical check-in data in standard supervised learning manner, which fail to fully explore each

user’s multi-faceted preferences, and suer from data scarcity and long-tailed POI distribution, resulting in sub-optimal performance.

To this end, we propose a

elf-

upervised

raph-enhanced POI

Rec

ommender (S

GRec) for next POI recommendation. In particular,

we devise a novel Graph-enhanced Self-attentive layer to incorporate the collaborative signals from both global transition graph and

and local trajectory graphs to uncover the transitional dependencies among POIs and capture a user’s temporal interests. In order to

counteract the scarcity and incompleteness of POI check-ins, we propose a novel self-supervised learning paradigm in S

GRec, where

the trajectory representations are contrastively learned from two augmented views on geolocations and temporal transitions. Extensive

experiments are conducted on three real-world LBSN datasets, demonstrating the eectiveness of our model against state-of-the-art

methods.

Additional Key Words and Phrases: Self-attention; Self-supervised Learning; Next POI Recommendation

1 INTRODUCTION

The rapid growth of Location-Based Social Networks (LBSNs), such as Yelp and Foursquare, has signicantly raised

attention on the studies of location-based recommendation in recent years. Among various location-based services,

next Point-of-Interest (POI) recommendation is the most prominent one since it can eectively assist service providers

to comprehensively understand user movement patterns for accurate customised promotions. Meanwhile, the next POI

recommenders also provide users with tailored trip plans, helping them to make decisions on the next move according

to their own travel history.

Previous next POI recommenders have been evolving from pure temporal sequential transitions by rst-order

Markov-chain [

] and tensor factorisation [

], to recurrent neural networks (RNNs) [

] and self-

attention mechanism [

] for long-term and short-term preference modelling. In addition to the exploitation of

Authors’ addresses: Yang Li, The University of Queensland, Brisbane, Queensland, Australia, 4068, yang.li@uq.edu.au; Tong Chen, The University

of Queensland, Brisbane, Queensland, Australia, 4068, tong.chen@uq.edu.au; Peng-Fei Zhang, The University of Queensland, Brisbane, Queensland,

Australia, 4068, mima.zpf@gmail.com; Zi Huang, The University of Queensland, Brisbane, Queensland, Australia, 4068, huang@itee.uq.edu.au; Hongzhi

Yin, The University of Queensland, Brisbane, Queensland, Australia, 4068, h.yin1@uq.edu.au.

Permission to make digital or hard copies of part or all 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 third-party

components of this work must be honored. For all other uses, contact the owner/author(s).

Manuscript submitted to ACM

Manuscript submitted to ACM 1

arXiv:2210.12506v1 [cs.LG] 22 Oct 2022

2 Yang Li, Tong Chen, Peng-Fei Zhang, Zi Huang, and Hongzhi Yin

Fig. 1. A overview of the proposed S2GRec framework.

sequential regularities, the RNN-based and self-attentive methods additionally explore the spatiotemporal eects of

geographical distances and time intervals between successive check-ins. However, the transition patterns among POIs

learned by these methods are limited to a

local

view, where only the co-occurrences of POIs within each independent

check-in sequence are captured. Intuitively, similar users sharing common preferences can be leveraged as a reference

for recommending next POI to a certain user. By converting check-in trajectories into graphs, recent graph neural

network (GNN)-based approaches [

] are able to take advantage of

global

spatial and temporal factors by taking

account of the correlated neighbour check-in sequences for recommendation. However, the traditional GNNs, such

as graph attention networks [

] and graph convolutional networks [

], suer from over-smoothing problem when

stacking more than 3 layers, which may fail to capture the long-range POI dependencies. In [

], we propose a new

GNN-based POI recommender that incorporates the category-wise correlation modelling to alleviate the negative

impacts from the sparse observed POI-POI interactions. However, SGRec still faces the challenges from the following

perspectives:

•

Missing check-in data: prior work purely rely on the observed check-in records for next POI recommendation,

where some possible check-ins between each successive check-in pairs are neglected since it is unlikely for a

user to check in on each visit. This results in incomplete check-in trajectories, which undermines the capability

of the existing work on POI correlation modelling.

•

Noisy check-in data: there exist many obvious unreal check-in records (e.g., two consecutive check-ins are far

apart in distance but occurred in a short time period) and incorrect location ID (e.g., a POI near the actual visiting

POI) in raw dataset due to the unstable GPS signal and mobile network signal, which impede the existing models

to reveal the actual POI transition correlations, resulting in sub-optimal performance.

Manuscript submitted to ACM

Self-supervised Graph-based Point-of-interest Recommendation 3

To address the aforementioned limitations and challenges, we make substantial technical advancements upon

SGRec [

], and propose a

elf-

upervised

raph-enhanced POI

Rec

ommendation framework (S

GRec). Figure 1

illustrates the overall workow of S

GRec. It rst derives POI representations from two constructed global graphs in

the pretraining stage, making each POI carries the knowledge of global spatial and temporal information. Then, to

address the long-range dependency modelling limitations of prior GNN-based approaches and take account of global

spatial and temporal factors for recommendation, we propose a novel

raph-enhanced

elf-

ttention

etwork (GSAN)

layer as the user preference encoder, which makes full use of the long-short-range dependency modelling capability

oered by the self-attention mechanism, and further aggregates the collaborative signals from the global graphs to

obtain the comprehensive trajectory representations. Finally, we address the two above-mentioned challenges through

an integration of self-supervision signals. Concretely, three carefully-designed augmentation operators are devised,

which are performed on the local trajectory graphs to mimic the real-life human check-in behaviours by studying the

underlying semantics among POIs, yielding dierent views of each training instance (i.e., the augmented trajectory

graphs). An auxiliary self-supervision task of optimising the consistency between the views of the same instance as well

as discriminating it from other training samples is introduced to boost the robustness of S

GRec to the data sparsity

and noise.

Our main contributions of this paper are summarised in the followings:

•

We propose a novel Self-supervised Graph-enhanced POI Recommender (S

GRec) based on self-attention

network to model the complex POI-POI relationships and make fully eective use of global spatial and temporal

information for next POI recommendation.

•

We propose an innovative graph-enhanced self-attentive user preference encoder, GSAN, which adequately

captures both of the complex high-order underlying POI-wise and category-wise dependencies from various

graph signals, while maintains the sequential properties within the constructed trajectory graphs.

•

We introduce a novel auxiliary self-supervision task that complements S

GRec’s capability of learning user

behaviours from sparse and noisy data by eective data augmentation with semantically-correlated POIs.

•

Extensive experiments are conducted on three real-world datasets demonstrating the ecacy of our framework

against several state-of-the-art next POI recommenders, and a detailed ablation study is launched on the

eectiveness of each proposed component.

2 RELATED WORK

In this section, we highlight the most recent work in next POI recommendation and contrastive self-supervised learning

in recommender systems, which are the most relevant to our work.

2.1 Next POI Recommendation

Dierent from the conventional POI recommendation that mainly captures user-POI relationships and geographical

impacts from users’ check-in activities, next POI recommendation is more challenging since it needs to pay additional

attention to the sequential patterns for the user’s next move prediction. Intuitively, the chronologically check-in

records carry strong spatio-temporal characteristics. Thus, various studies have been conducted to model the sequential

dependencies and spatial relationships among POIs for next POI recommendation. In general, the existing next POI

recommenders can be categorised into four main streams according to the employed techniques: tensor factorisation-

based, RNN-based, self-attention-based and GNN-based. The rst stream of next POI recommenders mainly come from

Manuscript submitted to ACM

4 Yang Li, Tong Chen, Peng-Fei Zhang, Zi Huang, and Hongzhi Yin

the early studies on user sequential behaviour modelling, which are primarily built upon tensor factorisation, where the

relationships among user, POI and time are exploited. By factorising the tensor of POI-POI transition cube, FPMC-LR [

]

predicts the user’s next move based on the exploited POI transition dependencies with location constraints. STELLER

[

] extends FPMC-LR by additionally introducing a POI-time tensor to capture the user’s periodical patterns. PRME

[

] proposes a ranking metric learning framework that learns individual preference and POI sequential dependencies in

multiple Euclidean spaces improving the recommendation performance. However, this stream of methods are developed

based on the assumption that all the factor components are independent and linearly combined, which neglect modelling

of these factor interactions.

With the great success of RNNs in sequential data modelling including NLP [

] and time-series prediction

[

], RNNs are widely adopted in the task of next POI recommendation. ST-RNN [

] denes two distance-specic

and time-specic projection matrices in RNN for spatio-temporal modelling. Similarly, Time-LSTM [

] and STGN [

]

enhance the capability of vanilla LSTM model by equipping spatio-temporal gates. LSTPM [

] exploits the long-term

user preferences by identifying anitive trajectories with a non-local operation. Inspired by the recent advances of

self-attention mechanism in sequential recommendation [

], two state-of-the-art self-attentive models are proposed.

GeoSAN [

] discretise the geographical information by a hierarchical gridding module performed on the map with a

self-attentive encoder, while STAN [

] employs linear interpolation to translate the continuous spatial distance and time

information into latent embeddings, and adopts a bi-directional self-attentive network for long-short-term preference

encoding. However, self-attentive models make predictions only conditioned on individual trajectories, which ignores

the useful transition information from other correlated trajectories. To this end, SGRec [

] and STP-UDGAT [

] are

proposed, which both construct a global graph from the all observable trajectories, and then employ graph attention

network (GAT) to capture the collaborative information across trajectories. Nevertheless, stacking too many GNN layers

will result in over-smoothing problem [

], which means GNN models can only attend on close neighbours, while the

long-range POI dependencies cannot be fully captured. In contrast, S

GRec takes advantages of both self-attention

architecture for long-short-term preference modelling and GNN for collaborative information exploitation by encoding

various contextual information from graph signals, while alleviates the mentioned limitations of existing methods.

2.2 Self-supervised Learning in Recommender Systems

In recent years, self-supervised learning (SSL) has achieved remarkable success in various areas ranging from NLP

[

], graph learning [

] and computer vision [

]. It constructs implicit supervisory signals from

unlabelled data helping the recommenders to learn the subtle patterns from the input data. In light of the promising

discriminative ability oered by SSL, it has been an emerging paradigm in recent recommender systems. S

-Rec [

]

applies random masking on items and item attributes, thus maximising the mutual information over the sequences and

item attributes. Yao et al. [

] propose to perform feature-level dropout and masking on items and item-category for

large-scale item recommendation. SGL [

] perturbs the user-item graph structure (i.e., node/edge dropout and random

walk) to generate multiple views of the same node, enabling SSL for item recommendation. CLS4Rec [

] develop

random augmentation operators launched on item sequences creating dierent views of sequence-level representations

for SSL. To the best of our knowledge, there is no existing work on exploring the self-supervision signals in next POI

recommendation. In S

GRec, we not only introduce random augmentation operators, which are similar to SGL [

]

but also devise a global context-aware augmentation operator that explores the POI spatial and temporal correlations

from the global view to create high-quality augmented views of the same trajectory graph, allowing the integration of

self-supervised learning scheme.

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：

Self-supervisedGraph-basedPoint-of-interestRecommendationYANGLI,TheUniversityofQueensland,AustraliaTONGCHEN,TheUniversityofQueensland,AustraliaPENG-FEIZHANG,TheUniversityofQueensland,AustraliaZIHUANG,TheUniversityofQueensland,AustraliaHONGZHIYIN,TheUniversityofQueensland,AustraliaTheexponentialgrowt...

展开>> 收起<<

Self-supervised Graph-based Point-of-interest Recommendation.pdf

共19页,预览4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！

Self-supervised Graph-based Point-of-interest Recommendation

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: