MAUIL: Multi-level Attribute Embedding for Semi-supervised User Identity Linkage

TL;DR

This paper introduces MAUIL, a semi-supervised model that leverages multi-level attribute embeddings and RCCA to improve user identity linkage across social networks, capturing semantic features at multiple levels.

Contribution

The paper proposes a novel multi-level attribute embedding approach combined with RCCA for semi-supervised user identity linkage, addressing limitations of single-method attribute representations.

Findings

MAUIL outperforms existing methods on real-world datasets.

Multi-level attribute features improve linkage accuracy.

The model effectively captures semantic information at character, word, and topic levels.

Abstract

User identity linkage (UIL) across social networks has recently attracted an increasing amount of attention due to its significant research challenges and practical value. Most of the existing methods use a single method to express different types of attribute features. However, the simplex pattern can neither cover the entire set of different attribute features nor capture the higher-level semantic features in the attribute text. This paper establishes a novel semisupervised model, namely the multilevel attribute embedding for semisupervised user identity linkage (MAUIL), to seek the common user identity across social networks. MAUIL includes two components: multilevel attribute embedding and regularized canonical correlation analysis (RCCA)-based linear projection. Specifically, the text attributes for each network are first divided into three types: character-level, word-level, and topic-level attributes. Second, unsupervised approaches are employed to extract the corresponding three types of text attribute features, and user relationships are embedded as a complimentary feature. All the resultant features are combined to form the final representation of each user. Finally, target social networks are projected into a common correlated space by RCCA with the help of a small number of prematched user pairs. We demonstrate the superiority of the proposed method over state-of-the-art methods through extensive experiments on two real-world datasets.