UTCS: Effective Unsupervised Temporal Community Search with Pre-training of Temporal Dynamics and Subgraph Knowledge Conference: Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval; July 13–18, 2025; Padua, Italy Proceedings of the 48th Int

Yue Zhang, Yankai Chen, Yingli Zhou, Yucan Guo, Xiaolin Han, Chenhao Ma · 2025

Read UTCS: Effective Unsupervised Temporal Community Search with Pre-training of Temporal Dynamics and Subgraph Knowledge Conference: Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval; July 13–18, 2025; Padua, Italy Proceedings of the 48th Int by Yue Zhang, Yankai Chen, Yingli Zhou, Yucan Guo, Xiaolin Han, Chenhao Ma (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-02784.txt.

Introduction

Yue Zhang and colleagues tackled community search in changing networks in a 2025 paper. In a temporal graph, links carry time stamps, and a community search asks who belongs with a given starting person or group. Older methods need the target shape in advance, while learning methods often ignore when links happened.

Their system first pre-trains on time patterns and small subgraph facts with no labels at all. At search time it gathers a candidate group, scores members with the learned features, and trims to a final community. Tests reported in the work show the unsupervised approach finding good groups without hand-set shapes. Time-aware search helps in social, trade, and contact networks alike.

Catalog record

CREATORYue Zhang, Yankai Chen, Yingli Zhou, Yucan Guo, Xiaolin Han, Chenhao Ma
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE23 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-02784

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

UTCS: Effective Unsupervised Temporal Community Search with Pre-training of Temporal Dynamics and Subgraph Knowledge Conference: Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval; July 13–18, 2025; Padua, Italy Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’25), July 13–18, 2025, Padua, Italy DOI: 10.1145/3726302.3730263 ISBN: 979-8-4007-1592-1/2025/07 CCS: Computing methodologies Neural networks CCS: Mathematics of computing Graph algorithms Yue Zhang, Yankai Chen, Yingli Zhou, Yucan Guo, Xiaolin Han, Chenhao Ma Abstract . In many real-world applications, the evolving relationships between entities can be modeled as temporal graphs, where each edge has a timestamp representing the interaction time. As a fundamental problem in graph analysis, community search (CS) in temporal graphs has received growing attention but exhibits two major limitations: (1) Traditional methods typically require predefined subgraph structures, which are not always known in advance. (2) Learning-based methods struggle to capture temporal interaction information. To fill this research gap, in this paper, we propose an effective U nsupervised T emporal C ommunity S earch with pre-training of temporal dynamics and subgraph knowledge model ( UTCS ). UTCS contains two key stages: offline pre-training and online search. In the first stage, we introduce multiple learning objectives to facilitate the pre-training process in the unsupervised learning setting. In the second stage, we identify a candidate subgraph and compute community scores using the pre-trained node representations and a novel scoring mechanism to determine the final community members. Experiments on five real-world datasets demonstrate the effectiveness. Figure 1 . The architecture of our model. 1. Introduction Graphs are widely utilized in various real-life fields ( Zeng et al. 2024b ; Zeng et al. 2024a ; Luo et al. 2023 ; Luo et al. 2024 ) , such as social networks, biological graphs, and financial graphs. Community Search (CS), as an important graph analytical problem, has attracted significant attention recently ( Fang et al. 2019 ) . Given a graph and a set of query nodes, the community search problem aims to find a cohesive and dense subgraph containing the query nodes ( Sozio and Gionis 2010 ) . CS is widely used in applications like social recommendation, research community detection, and fraud group identification, inspiring numerous specialized approaches. Existing methods can be broadly classified into two categories: traditional CS algorithms ( Akbas and Zhao 2017 ; Yuan et al. 2017 ; Fang et al. 2017 ; Chen et al. 2018 ; Zhou et al. 2023 ; Zhou et al. 2024a ) and learning-based CS models ( Gao et al. 2021 ; Jiang et al. 2022 ; Hashemi et al. 2023 ; Wang et al. 2024a ; Wang et al. 2024b ) . Traditional CS algorithms rely on predefined subgraph patterns, such as k -core ( Cui et al. 2014 ; Sozio and Gionis 2010 ; Chen et al. 2021 ) , k -truss ( Akbas and Zhao 2017 ; Huang et al. 2015 ) , and k -clique ( Cui et al. 2013 ; Yuan et al. 2017 ; Zhou et al. 2024b ) . These models impose strict topological constraints on the community, which may not always align with the structures of real-world communities. For instance, the k -core-based models assume that each node within the community has a degree of at least k , which may not always hold in reality. In contrast, learning-based CS models like QD-GNN ( Jiang et al. 2022 ) and TransZero ( Wang et al. 2024a ) adapt flexibly to graph structures without predefined patterns: QD-GNN separately models queries and communities via GNNs, while TransZero uses Graph Transformers to capture node-community proximity. Despite the success of learning-based CS studies, most are designed for static graphs and overlook crucial temporal interaction information . In real-world temporal graphs ( Chu et al. 2019 ) , such as business collaborations, two parties may establish relationships during specific periods. Identifying communities in such graphs is crucial for many real-world applications and has garnered growing research attention. While temporal graphs can be treated as static for input into static models, this often leads to sub-optimal results. For example, a recent learning-based work ( Hashemi et al. 2023 ) considers the graph within each time window as static, failing to capture temporal interactions within the time window. Moreover, most learning-based methods heavily rely on costly labeled data. This highlights the urgent need for an unsupervised approach that captures temporal information for effective community search. To date, there is no existing learning-based approach for temporal community search in the unsupervised setting . Our solution. To address the aforementioned limitations, we have developed an effective model, namely U nsupervised T emporal C ommunity S earch with pre-training of temporal dynamics and subgraph knowledge ( UTCS ). UTCS incorporates temporal and structural information at both local and global scales. As a learning-based method, it operates without requiring a predefined subgraph structure. Our model has two phases: offline pre-training and online search. During pre-training, we introduce the multiple learning objectives through three components: a temporal dynamics learner based on the Hawkes process to model interaction patterns, a subgraph alignment mechanism leveraging Student’s t-distribution and KL divergence to capture structural dependencies, and a batch-wise embedding refinement to reconstruct adjacency relations. This phase enables unsupervised and label-free training. In the online search phase, we first locate candidate subgraphs, then leverage learned node representations to compute community scores and employ a dedicated matching mechanism to predict the final community structure. Our main contributions are as follows. • We propose a learning-based model for the temporal CS problem and, to the best of our knowledge, this is the first unsupervised approach to tackle it. • Our model follows a two-step framework with modules designed to capture temporal, local, and global structural information. Additionally, it employs a local search approach to significantly reduce time and space complexity. • Extensive experiments on five real-world datasets demonstrate the effectiveness of the model, achieving an average improvement of 60.44% in F1-score compared to the latest competitors. 2. Problem Formulation In this paper, we consider an undirected temporal graph \mathcal{G(V,E)} , where \mathcal{V} and \mathcal{E} denote the sets of nodes and edges, respectively. Let |\mathcal{V}|=n and |\mathcal{E}|=m . Each edge e\in\mathcal{E} is a triplet (u,v,t) with u,v\in\mathcal{V} and t\in\mathbb{N} representing the interaction timestamp between u and v . Note that (u,v,t_{1}) and (u,v,t_{2}) are considered as two edges when t_{1}\neq t_{2} . We introduce the concept of temporal neighbors N_{(u,t)} , where N_{(u,t)}=\{w|(u,w,t_{w})\in\mathcal{E}\text{ and }t_{w}<t\} . Figure 1 preliminaries illustrate a sample temporal graph \mathcal{G} with 6 vertices and 17 temporal edges. In addition, we define the concept of the de-temporal graph, a version of the temporal graph \mathcal{G} with timestamps ignored as \mathcal{G^{\prime}(V^{\prime},E^{\prime})} , where \mathcal{E}^{\prime}=\{(u,v)|\exists(u,v,t)\in\mathcal{E}\} and |\mathcal{E}^{\prime}|=\bar{m} . Figure 1 preliminaries show a de-temporal graph \mathcal{G^{\prime}} . Problem 1 (Temporal Community Search (TCS) ( Hashemi et al. 2023 ) ). Given a temporal graph \mathcal{G} and a set of query nodes \mathcal{Q}\subseteq\mathcal{V} , the TCS problem aims to return a subgraph (i.e., community) \mathcal{T}_{\mathcal{Q}} containing \mathcal{Q} . In this paper, we focus on the learning-based approach to solve the TCS problem. For a set of query nodes \mathcal{Q}\subseteq\mathcal{V} , and its

Read the complete text