TL;DR
This paper introduces a novel "Multi-grained Alignment Method" (MGA) to tackle the challenge of accurately aligning users and communities across different social networks. By focusing on "stable topics" derived from user-generated content and employing a hierarchical, multi-level community detection approach, MGA significantly improves alignment accuracy compared to existing methods, especially for users on the fringes of communities.
What Problem Does This Paper Address?
In today's interconnected world, users often maintain profiles across multiple social networks. Identifying and linking these cross-network identities (user alignment) is crucial for enhancing user experience, improving recommendation systems, and enabling more comprehensive social analysis. However, existing methods face significant hurdles. User alignment struggles with privacy concerns and the difficulty of obtaining comprehensive user features. Group alignment, while less constrained by privacy, suffers from low accuracy due to the sheer volume of "edge users" (users with few connections) and often relies on fixed, predefined community structures. This paper aims to bridge this gap by proposing a method that can align users and communities at various granularities, from broad community structures down to individual users, while effectively handling the complexities of real-world social networks.
Methods
The MGA method is built upon several key innovations:
- Stable Topic Acquisition: Instead of relying on fleeting trends, MGA first identifies "stable topics" from user-generated content (UGC). This involves clustering keywords, calculating a "jitter degree" based on temporal activity, and filtering out short-lived topics to focus on enduring user interests. These stable topics are then embedded into a vector space using Word2Vec.
- Stable Topic-Louvain (ST-L) for Multi-Level Community Detection: An enhanced Louvain algorithm is introduced, named ST-L. This algorithm uses the calculated user topic similarity to dynamically update edge weights in the social network. This allows for hierarchical community detection without requiring pre-defined labels, enabling the discovery of communities at multiple levels of granularity.
- Fuzzy Center-Based Clustering for Community Feature Extraction: To capture the nuanced thematic features of communities, an improved fuzzy center-based clustering algorithm is employed. This method optimizes the correlation between user theme vectors and community feature points, leading to more accurate representation of community themes.
- Multi-Granularity Iterative Alignment: The core of MGA is its iterative alignment process. It starts with coarse-grained community alignment and progressively refines it to finer-grained sub-communities, eventually reaching user-level alignment. This multi-layered approach allows for alignment at any desired level, mitigating the accuracy issues associated with single-granularity methods.
Datasets
The paper utilizes real-world datasets for its experiments:
- Weibo-Zhihu: A custom dataset collected by crawling Weibo and Zhihu. Users with fewer than 20 posts per year were filtered out to ensure topic relevance. Positive samples were obtained by identifying Weibo accounts and high-impact real-name authenticated users in Zhihu. Negative samples were randomly captured. The dataset includes 5,876 users, 61,002 contacts, and 138,749 posts.
- Aminer and Cora: Publicly available, open-labeled datasets used for evaluating community detection performance. Aminer contains 12,840 users and 190,658 edges across four categories, while Cora has 2,708 users and 5,278 edges across seven categories.
Key Insights and Ideas
- The Power of "Stable Topics": Focusing on enduring user interests rather than transient trends is crucial for robust cross-network alignment. This approach provides a more stable foundation for understanding user behavior and community structure.
- Hierarchical Alignment is Key: Aligning communities and users at multiple levels of granularity, from broad groups to individual users, is essential for overcoming the limitations of single-level approaches. This allows for a more nuanced and accurate representation of relationships.
- Fuzzy Logic for Edge Users: Employing fuzzy classification techniques helps to better incorporate "edge users" (those with fewer connections) into the alignment process, preventing them from disproportionately impacting accuracy.
- Dynamic Community Structures: The ST-L algorithm's ability to detect communities without pre-set labels is a significant advantage, allowing it to adapt to the organic and evolving nature of social networks.
Key Results
- Community Alignment: MGA demonstrated improved accuracy in community alignment compared to CAlign (6.1% improvement) and PERFECT (2.5% improvement) on the microbot-Zhihu dataset.
- User Alignment: On the Weibo-Zhihu dataset, MGA achieved superior precision, recall, and F1 scores compared to several mainstream user alignment algorithms, including BSNA, PERFECT, UGCLink, MEgo2Vec, and MUIUI. Specifically, MGA improved accuracy by 4.5% over MEgo2Vec and 3.4% over MUIUI, with corresponding improvements in recall and F1 values.
- Community Detection Performance: The ST-L algorithm, a component of MGA, showed strong performance in community detection, outperforming single-layer methods and competitive with other hierarchical methods, particularly on datasets with pre-defined classifications.
Limitations and Future Work
- Future Work on Edge Weights: The authors suggest integrating stable topic-based edge weights more deeply into user structure and attribute features to create more diverse weights and further enhance hierarchical community detection.
- Integrating More Features: Future work could involve further improving the multi-granularity alignment method by integrating structure, attributes, and content more comprehensively. This would aim to reduce the impact of edge users on community alignment and boost overall accuracy.
- Computational Complexity: While not explicitly detailed as a limitation, the iterative nature of the multi-grained alignment and community detection processes suggests potential computational demands, especially on very large networks.
Summary updated