Mining Unfollow Behavior in Large-Scale Online Social Networks via Spatial-Temporal Interaction

Haozhe Wu, Zhiyuan Hu, Jia Jia, Yaohua Bu, Xiangnan He, Tat-Seng Chua

Abstract

While social network research has traditionally focused on how users connect, the dissolution of social ties—unfollowing—remains relatively unexplored despite its frequency in digital life. This paper introduces the Unified Model with Heterogeneous Information (UMHI), a deep learning framework designed to predict unfollow behavior by modeling the non-linear interactions between a user's spatial network position and their temporal activity. Using a massive Sina Weibo dataset, the study demonstrates that UMHI outperforms traditional methods by over 16% in precision, highlighting the critical role of spatial-temporal interaction in understanding network evolution.

Introduction

In Online Social Networks (OSNs), relationships are inherently dynamic. While "link prediction" (identifying future connections) is a well-studied field, "link dissolution" (predicting when a tie will break) is equally critical for understanding network health. An unfollow decision is rarely a reaction to a single event; rather, it is a complex interplay of spatial factors (a user's role and position in the network), temporal factors (recent posting activity and unfollow history), and the non-linear interaction between them. For instance, a user might tolerate low-quality content from a close friend but unfollow a distant acquaintance for the same behavior.

Methodology: The UMHI Framework

To address these complexities, the authors developed the UMHI (Unified Model with Heterogeneous Information) framework. As shown in Figure 2, the model integrates diverse data sources through specialized encoders:

  • Network Structure Encoder (Spatial): Utilizes the LINE algorithm to capture local network neighborhoods, representing users as vectors based on 1st and 2nd order proximity.
  • Post Content Encoder (Temporal): Employs a Hierarchical Attention Network (HAN) to weigh the importance of specific words and posts within a user's timeline.
  • Unfollow History Encoder (Temporal): Applies Matrix Factorization to an "unfollow co-rating matrix" to identify latent patterns in user behavior.
The UMHI framework architecture.

The outputs of these encoders are processed through Multi-Layer Perceptron (MLP) layers. This fusion strategy allows the model to learn non-linear interactions, such as how a user's social role (spatial) modulates the impact of their content (temporal) on their followers' decisions.

Dataset and Empirical Observations

The study utilizes a massive benchmark dataset from Sina Weibo, featuring 1.8 million users and 400 million social relationships. Over a one-month observation period, the researchers tracked 10 million unfollow records. Notably, 40% of users unfollowed at least one person during this time, confirming that link dissolution is a pervasive phenomenon in modern OSNs.

Key Insights

The researchers identified several critical patterns in unfollow behavior:

  1. Role-Based Stability: "Opinion leaders" (high PageRank) generally exhibit more stable relationships. However, their followers are significantly more sensitive to content similarity.
  2. Structural Bridges: Users acting as "structural holes"—connecting disparate communities—are less affected by content similarity, as their value is derived from their network position.
  3. Predictive Power of History: Latent patterns in unfollow history were found to be stronger predictors of future unfollows than the actual text content of posts.
Unfollow ratios vs. content similarity and exposure.

As illustrated in Figure 4, the ratio of unfollow actions decreases as content similarity increases, particularly for opinion leaders. Conversely, the relationship with "exposure" (number of posts) shows that excessive activity can lead to higher unfollow rates for ordinary users, while opinion leaders maintain a more stable buffer.

Results and Conclusion

UMHI achieved a precision of 0.7868, representing a 16.44% improvement over baseline methods like Doc2vec combined with Logistic Regression. Ablation studies confirmed that the integration of spatial and temporal data is essential for high performance.

While the model is currently conservative (lower recall due to the sparsity of unfollow events), it demonstrates high robustness even with limited training data. Future work aims to apply these insights to anomaly detection, potentially serving as an early warning system for social dissolution or the spread of misinformation.

Summary updated

BibTeX
@inproceedings{wu2020mining,
  author = {Haozhe Wu and Zhiyuan Hu and Jia Jia and Yaohua Bu and Xiangnan He and Tat-Seng Chua},
  title = {{Mining Unfollow Behavior in Large-Scale Online Social Networks via Spatial-Temporal Interaction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i01.5358},
}
Figure