GNN Applied to Ego-nets for Friend Suggestions
Evgeny Zamyatin·December 16, 2024
Summary
The Generalized Ego-network Friendship Score framework uses WalkGNN for friend suggestion in social networks, reducing link prediction to low-scale ego-net tasks. This scalable approach aggregates results and excels in heterogeneous, dynamic graph-level link prediction. Evaluated on the Ego-VK dataset, it outperforms baselines, and A/B tests show improved business metrics. A distributed triangle counting algorithm computes common neighbors and Adamic-Adar heuristics in large graphs, making ego-nets a compact way to analyze local neighborhoods in graphs with billions of nodes. Inspired by friend suggestion algorithms, WalkGNN, a second-order GNN, constructs representations for node pairs, considering different link types and numerical characteristics. Evaluated on the Ego-VK dataset from Russia's largest social network, VK, this model efficiently solves link prediction tasks in complex networks.
Introduction
Background
Overview of social network analysis and link prediction
Importance of friend suggestion algorithms in social networks
Challenges in link prediction for large-scale, heterogeneous, and dynamic graphs
Objective
To present a scalable framework for link prediction using the Generalized Ego-network Friendship Score
To introduce WalkGNN, a second-order Graph Neural Network (GNN) for efficient link prediction
Method
Data Collection
Description of the Ego-VK dataset used for evaluation
Data sources and preprocessing steps for the dataset
Data Preprocessing
Techniques for handling large-scale, heterogeneous, and dynamic graph data
Methods for aggregating and normalizing data for model training
Model Architecture
Detailed explanation of WalkGNN architecture
Incorporation of second-order information in node representations
Handling of different link types and numerical characteristics
Evaluation
Metrics used for assessing the performance of the Generalized Ego-network Friendship Score framework
Comparison with baseline models on the Ego-VK dataset
Results
Presentation of experimental results on the Ego-VK dataset
Analysis of improvements in business metrics from A/B tests
Scalability and Efficiency
Distributed Triangle Counting Algorithm
Overview of the algorithm for computing common neighbors and Adamic-Adar heuristics
Explanation of how it handles large graphs with billions of nodes
Ego-net Analysis
Explanation of how ego-nets provide a compact way to analyze local neighborhoods
Benefits of using ego-nets for scalable link prediction in complex networks
Conclusion
Summary of the Generalized Ego-network Friendship Score framework
Future Work
Potential improvements and extensions of the framework
Areas for further research in scalable link prediction and social network analysis
Basic info
papers
social and information networks
artificial intelligence
Advanced features
Insights
Which dataset was used to evaluate the performance of the Generalized Ego-network Friendship Score framework, and what were the results compared to baseline models?
How does the framework utilize WalkGNN for friend suggestion in social networks?
What are the key advantages of this scalable approach in heterogeneous, dynamic graph-level link prediction?
What is the main idea behind the Generalized Ego-network Friendship Score framework?