Scholarlink

<p>Accurate user representation, which learns meaningful embeddings from interaction data, is crucial for applications such as personalized recommendation, fraud detection, and behavioral analysis. Existing surveys typically categorize user representation learning models from three perspectives: model-centric, task-centric, and architecture-centric. Complementary to these perspectives, this survey introduces a <b>data-centric view</b>, which focuses on the intrinsic properties of user interaction data and establishes a principled connection between data properties and model assumptions. From this data-centric perspective, we classify user representation learning models according to two fundamental aspects of interaction data: <b>topological</b> and <b>temporal</b> properties. Accordingly, three key paradigms are identified: (1) <b>static interaction graphs</b>, which focus on topological relationships; (2) <b>interaction sequences</b>, which model temporal dynamics; and (3) <b>dynamic interaction graphs</b>, which integrate both topological relationships and temporal dynamics. Additionally, we explore enhanced semantic and multimodal information integration to enhance representational power. This extension leads to two emerging paradigms: (4) <b>heterogeneous interaction graphs</b> and (5) <b>multimodal interaction graphs</b>. Finally, we review benchmark datasets and outline open challenges, including structure-aware integration with large language models (LLMs) and scalable learning in multi-source environments. Overall, this survey analyzes the relationship between model design and intrinsic data properties across various paradigms, highlighting how data properties inform the selection of appropriate modeling approaches and help anticipate their limitations in diverse application scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Exploring user representation learning from interaction data: a data-centric survey

  • 期刊论文

  • 发表日期：

    2026年3月17日

查看全文
  • Shenghai Zhong,
  • Shu Guo,
  • Chang Liu

摘要

Accurate user representation, which learns meaningful embeddings from interaction data, is crucial for applications such as personalized recommendation, fraud detection, and behavioral analysis. Existing surveys typically categorize user representation learning models from three perspectives: model-centric, task-centric, and architecture-centric. Complementary to these perspectives, this survey introduces a data-centric view, which focuses on the intrinsic properties of user interaction data and establishes a principled connection between data properties and model assumptions. From this data-centric perspective, we classify user representation learning models according to two fundamental aspects of interaction data: topological and temporal properties. Accordingly, three key paradigms are identified: (1) static interaction graphs, which focus on topological relationships; (2) interaction sequences, which model temporal dynamics; and (3) dynamic interaction graphs, which integrate both topological relationships and temporal dynamics. Additionally, we explore enhanced semantic and multimodal information integration to enhance representational power. This extension leads to two emerging paradigms: (4) heterogeneous interaction graphs and (5) multimodal interaction graphs. Finally, we review benchmark datasets and outline open challenges, including structure-aware integration with large language models (LLMs) and scalable learning in multi-source environments. Overall, this survey analyzes the relationship between model design and intrinsic data properties across various paradigms, highlighting how data properties inform the selection of appropriate modeling approaches and help anticipate their limitations in diverse application scenarios.