Computational Intelligence Methods for User Matching

Abstract

User identification refers to the process of user matching accounts across various social media platforms, which has numerous real-world applications. However, there are still many issues here, mainly in efficiency and effectiveness. As the time complexity of direct one-to-one user matching is O(mn) (Suppose there are m users on one platform and n users on another platform), the computation time increases exponentially as the number of users grows. Therefore, we explored methods to reduce the number of matching pairs. Before beginning formal computation, we propose method to filter users by record data, thereby eliminating the vast majority of unlikely candidate pairs and retaining as many real candidate pairs as possible. This approach can significantly reduce computation time. Besides, current user trajectory-based methods tend to focus separately on spatial and temporal data and fail to fully leverage the interdependence between them. In contrast, our approach integrates spatial-temporal information to enhance user identification accuracy, through a three-step process. First, we use kernel density estimation to measure the similarity of users’ trajectories, taking both spatial and temporal information into account. Second, we assign weights to each check-in record to prioritize discriminative ones. Finally, we utilize inconsistencies among check-in records to compute penalties for trajectory similarity. By identifying account pairs with similarity scores above a predefined threshold, we can determine whether they belong to the same user. We evaluated our method on three ground-truth datasets, demonstrating its competitive performance.