DSINE: Deep Structural Influence Learning via Network Embedding

Jianjun Wu, Ying Sha, Bo Jiang, Jianlong TanAAAI 2019 · DOI

Abstract

Social influence analysis is a cornerstone of modern network science, yet traditional network embedding methods often struggle with data sparsity and locality bias. This paper introduces DSINE (Deep Structural Influence Network Embedding), a deep learning framework designed to capture multi-scale structural information. By aligning a proximity-based auto-encoder with a "self-network" trained on random walks, DSINE effectively bridges the gap between local interactions and global network structures, providing a robust representation of user influence even in sparse datasets.

Introduction

In social network analysis, understanding how influence propagates is essential for applications ranging from viral marketing to personalized recommendation engines. Most existing network embedding techniques—which map graph nodes into low-dimensional vector spaces—face two significant hurdles. First, locality bias limits these models to immediate neighbors (first-order proximity), causing them to miss the broader macro-structures of influence. Second, data sparsity is endemic to social networks; most users interact with only a fraction of the population, leading to sparse adjacency matrices that hinder effective learning.

To address these challenges, the authors propose DSINE. This model learns representations that simultaneously preserve micro-, meso-, and macro-scale structures by integrating direct proximity data with inferred structural patterns.

Methodology

DSINE utilizes a dual-stream architecture to synthesize different levels of network information. The framework is composed of two primary components that are trained jointly through an alignment process.

The Origin Proximity Network ($r$)

This stream employs an encoder-decoder structure to process feature vectors representing multiple levels of proximity: first-order (direct links), second-order (shared neighbors), and high-order (distant reach). To handle the inherent sparsity of social graphs, the model uses a weighted loss function. This ensures the reconstruction process prioritizes observed interactions over the vast number of non-existent links, preventing the model from being biased toward "zeros."

The Aligned Self-Network ($o$)

To supplement the sparse proximity data, DSINE introduces a parallel "self-network." This component learns embeddings based on random walks guided by prior knowledge. These walks effectively "dense up" the network by exploring potential influence paths that may not be explicitly recorded as direct links.

Information Fusion and Alignment

The core innovation of DSINE is the joint objective function that forces the embeddings from the origin network and the self-network to align. This fusion allows the richer signal from the random walks to fill the gaps in the proximity data, resulting in a more comprehensive representation of a node's structural role.

Architecture of the DSINE framework showing the dual-stream encoder-decoder and self-network alignment.

Experiments and Results

The effectiveness of DSINE was evaluated on two real-world social media datasets: Twitter (crawled data) and RepLab-2014 (a benchmark for reputation monitoring). The model was compared against established baselines including DeepWalk, LINE, SDNE, and Node2vec in a multi-label classification task.

Performance Metrics

As shown in the experimental results, DSINE consistently outperformed all baseline models across various training data ratios. On the Twitter dataset with 90% training data, DSINE achieved a Micro-F1 score of 73.08%, a significant margin over Node2vec (59.0%) and SDNE (64.36%).

Performance comparison of DSINE against baseline algorithms across different training data percentages.

Key Findings

  • Robustness to Sparsity: Even when training data was restricted to 10%, DSINE maintained a competitive lead, demonstrating the efficacy of its alignment strategy in extracting signal from limited information.
  • Macro-F1 Gains: The model showed dramatic improvements in Macro-F1 scores (reaching 78.52% on Twitter), indicating a superior ability to identify niche or minority influence categories that other models often overlook.
  • Structural Equivalence: By incorporating high-order proximity, DSINE captures "structural equivalence," identifying users who play similar roles in the network even if they are not directly connected.

Discussion and Conclusion

DSINE represents a significant step forward in structural influence learning. By treating sparsity as a challenge to be solved through information fusion rather than a limitation to be ignored, the model provides a more nuanced view of social influence.

While the current iteration focuses heavily on user-user relations, future work could expand the framework to handle more complex heterogeneous networks (e.g., distinguishing between likes, retweets, and mentions). Additionally, applying these structural embeddings to large-scale community detection remains a promising avenue for further research. Despite the potential computational overhead of the dual-network alignment, DSINE offers a powerful tool for understanding the complex, multi-layered nature of influence in modern social graphs.

Summary updated

BibTeX
@inproceedings{wu2019dsine,
  author = {Jianjun Wu and Ying Sha and Bo Jiang and Jianlong Tan},
  title = {{DSINE: Deep Structural Influence Learning via Network Embedding}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.330110065},
}
Figure