DeepLink: A Deep Learning Approach for User Identity Linkage

1. Introduction

Feature-based approaches 基于特征的方法

从用户资料和社交平台的活动中提取独立的特征：e.g., username, gender, writing style

Network-based approaches 基于网络的方法

COSNET

IONE

Challenge

mapping the behaviors of cross-site accounts to a particular user

user representation：

• cross-platform behavior prediction
• cross-platform account correlation

lack of labeled data

Contributions

• 应用深度学习学习用户活动和网络结构的潜在语义
• semi-supervised graph regularization to predict the context (neighboring structures) of
  nodes in the network
• dual-learning process

2. Related Work

Profile-based methods

使用用户的用户资料信息，将不同社交平台的账号链接到一起

user name

writing style identification：grammatical structure，frequency of letters

Network-based methods

BIG-ALIGN

Neighborhood-based features

CLF

将网络节点的结构进行嵌入

3. Preliminary Background

Problem Definition

Social Network Graph

Network Embedding Model

User Identity Linkage

Graph Mapping Function

4. DeepLink：The Proposed Model

Network Structure Sampling

首先通过几轮随机游走生成每一个user的多个 social sequence （每一个 social sequence 代表了用户的社交关系）

从顶点(用户) ui 开始，在每一步中沿着随机选择的边前进，直到到达长度L

algorithms：LINE，GraRep， node2vec

User Latent Space Embedding

Neural Mapping Learning

two Multi-Layer Perceptrons

Linkage Dual Learning

Unsupervised UIL Pretraining ：先从一个graph映射到另一个graph再映射回来

Supervised UIL Learning

$$r_{s,t}^{a} = \frac{1}{k}\sum^{k}_{i=1}log(cos(v(u_{i}),v^{'}(u_{a}))+1) (6) \\ r_{t,s}^{a} = \frac{1}{k}\sum^{k}_{i=1}log(cos(\Phi^{-1}(v(u_{i})),v(u_{a}))+1) (6)$$

5. Experiments

Dataset

Baseline and settings

比较方法：

• Input-Output Network Embedding (IONE) : node vector, input vector and output vector
• ONE: node vector and output vector
• MAH: Manifold Alignment on Hypergraph
• MAG: Manifold Alignment on traditional Graphs
• CRW: Collective Random Walk

Evaluation Metrics

Precision@k(P@k)

MAP

AUC

Hit-Precision

$$MAP = \frac{\sum^{n}\frac{1}{ra}}{n} \\ AUC = \frac{\sum^{n}\frac{m+1-ra}{m}}{n} \\ Hit-Precision = \frac{\sum^{n}\frac{m+2-ra}{m+1}}{n} \\ ra\;is\; the\; rank\; of\; the\; positive\;matching\;identity \\ m\;is\; the\; number\; of\;negative\;user\;identities$$