User Identity Linkage across Social Media via Attentive Time-aware User Modeling

[IEEE TMM 2021] Official implementation of UserNet, an attentive time-aware framework for user identity linkage across social media platforms.

Authors

Xiaolin Chen¹, Xuemeng Song¹*, Siwei Cui², Tian Gan¹, Zhiyong Cheng³, Liqiang Nie¹

¹ Shandong University, Shandong, China
² Texas A&M University, USA
³ Qilu University of Technology (Shandong Academy of Sciences), China
* Corresponding author

Links

• Paper: IEEE Xplore
• Testing Dataset: Google Drive
• Training Dataset & Feature: Baidu Netdisk (Password: yw9f)

---

Table of Contents

• Updates
• Introduction
• Highlights
• Method / Framework
• Project Structure
• Installation
• Dataset
• Usage
• Citation
• Acknowledgement
• License

---

Updates

• [07/2019] Initial release of UserNet implementation
• [11/2020] Paper accepted at IEEE Transactions on Multimedia (TMM)

---

Introduction

This repository is the official implementation of the paper "User Identity Linkage across Social Media via Attentive Time-aware User Modeling", published in IEEE Transactions on Multimedia (TMM), 2021.

User identity linkage, which aims to identify accounts belonging to the same real-world entity across different social networks, is an increasingly important task. Existing methods generally overlook the temporal correlations among a user's posts. To address this issue, we propose UserNet, an attentive time-aware user modeling framework. UserNet captures the temporal patterns of users' posting behaviors and models the semantic correlations among multi-modal posts with an attention mechanism, thereby learning more discriminative user representations for identity linkage. We also release a large-scale multimodal dataset collected from Twitter and Foursquare to facilitate future research.

---

Highlights

• Proposes an attentive time-aware user modeling approach for user identity linkage
• Models temporal correlations among a user's multi-modal posts for more discriminative representations
• Designs attention mechanisms to capture semantic correlations across different modalities
• Releases a large-scale multimodal dataset from Twitter and Foursquare
• Achieves state-of-the-art performance on user identity linkage benchmarks

---

Method / Framework

Figure 1. Overall framework of UserNet. The model captures temporal posting patterns and multi-modal semantic correlations via attentive time-aware user modeling.

---

Project Structure

.
├── assets/                # Figures and framework diagrams
├── Dataset                # Dataset description
├── Time-M2M-Thread.py     # Main entry point for training and evaluation
├── att_lstm.py            # Attention-based LSTM module
├── attention.py           # Attention mechanism
├── lstm.py                # LSTM module
├── similarity.py          # Similarity computation
├── framework.png          # Framework figure
├── README.md
└── ...

---

Installation

1. Clone the repository

git clone https://github.com/iLearn-Lab/UserIdentityLinkage-UserNet.git
cd UserIdentityLinkage-UserNet

2. Prerequisite

• Python 3.5.2
• TensorFlow 1.9.0
• NumPy 1.16.4

3. Install dependencies

pip install numpy==1.16.4 tensorflow==1.9.0

---

Dataset

We provide a large-scale multimodal dataset collected from Twitter and Foursquare for user identity linkage research.

• Testing Dataset: Google Drive
• Training Dataset & Feature: Baidu Netdisk (Password: yw9f)

After downloading, please place the data files in the project directory and ensure the paths are configured correctly.

---

Usage

Training and Evaluation

python Time-M2M-Thread.py

---

Citation

If you find this work useful for your research, please cite our paper:

@article{chen2021usernet,
  title={User Identity Linkage Across Social Media via Attentive Time-Aware User Modeling},
  author={Chen, Xiaolin and Song, Xuemeng and Cui, Siwei and Gan, Tian and Cheng, Zhiyong and Nie, Liqiang},
  journal={IEEE Transactions on Multimedia},
  volume={23},
  pages={3957--3967},
  year={2021},
}

---

Acknowledgement

• Thanks to our supervisor and collaborators for valuable support.
• Thanks to the open-source community for providing useful baselines and tools.

---

License

This project is released under the Apache License 2.0.