Abstract
With the rapid development of the Internet of Things (IoT), law enforcement agencies have deployed extensive sensor networks for public safety. These diverse trajectories provide an unprecedented opportunity for intelligence analysis. While geographic co-movement patterns are prevalent in other fields, they have rarely been used to infer social connections in policing. Previous studies focused on singular trajectory forms with exact co-locations, often overlooking the spatial spread of these encounters. We propose Resonance, a novel framework for detecting co-occurrence patterns over multiplex trajectories. By constructing a "space-time resonance honeycomb" and introducing the "Geo-Spread" feature, our framework effectively distinguishes genuine social links from coincidental encounters. Experiments on public and real-world surveillance datasets demonstrate that Resonance outperforms state-of-the-art methods.
Introduction
In modern urban environments, "digital breadcrumbs" are scattered across various sensor networks, including traffic cameras, WiFi hotspots, and mobile base stations. Turning this multiplex trajectory data into an accurate map of social networks is a significant challenge for intelligence analysis. Existing methods typically suffer from two limitations: they focus on single-source data (data silos) and fail to distinguish between coincidental meetings at popular locations and intentional social interactions. The Resonance framework addresses these by mining co-occurrence patterns across diverse sensor types and introducing spatial metrics that capture the significance of shared movements.
The Resonance Framework
The Resonance framework operates through a multi-stage pipeline designed to transform raw sensor data into social connection probabilities.
Spatio-Temporal Resonance Honeycomb
Traditional methods often partition space into rigid grids, which fails to account for physical barriers like rivers or the actual distribution of sensors. Resonance utilizes a "space-time resonance honeycomb" constructed via DBSCAN clustering of sensor locations. This approach creates polygonal zones that reflect the true reachability and density of the sensor network, incorporating road network data and Regions of Interest (ROI).
Sliding Prism Recording
To capture co-occurrences accurately, the framework treats the time-axis as a third dimension. A "sliding prism" approach is employed, where a temporal window moves through the honeycomb zones to log whenever two or more objects appear in the same zone simultaneously. This method ensures that events occurring at the boundaries of fixed time windows are not lost.
The Geo-Spread Metric
The core innovation of Resonance is the Geo-Spread feature. It measures the standard deviation of the distances between all meeting points of a pair and their geometric center. The underlying hypothesis is that individuals with a genuine social bond are likely to meet across a variety of geographically dispersed locations. In contrast, strangers are more likely to "co-occur" repeatedly at a single high-traffic hub (e.g., a major transit station).
Experimental Evaluation
The framework was evaluated against three major datasets:
- Gowalla & Brightkite: Public location-based social network datasets containing millions of check-ins.
- Sichuan PSB Dataset: A real-world surveillance dataset provided by the Sichuan Public Security Bureau, featuring 1,057 labeled "vehicle-phone" trajectory pairs.
Key Results
Resonance consistently outperformed state-of-the-art methods like EBM, PGT, and SCI+. On public datasets, it achieved an average AUC of 0.9299. Most notably, on the real-world multiplex dataset from the Sichuan PSB, the framework reached an AUC of 0.9707, a substantial improvement over the 0.8451 achieved by the EBM method. The Geo-Spread feature proved to be the most powerful individual indicator, yielding an AUC of 0.9152 on its own in the surveillance context.
Conclusion
Resonance demonstrates that analyzing the geographic spread and context of co-occurrences across multiplex trajectories significantly improves the accuracy of social connection inference. While the computational cost of certain features remains a challenge for massive datasets, the framework provides a robust foundation for future intelligence analysis tools. Future work will explore integrating non-spatial data, such as financial transactions, to further enhance predictive capabilities.
Summary updated