Adaptive Geospatial Data Representation for Data Analytics and Sharing in the Mobility Domain

Abstract

In recent years, the mobility domain has gained attention from urban planners and researchers due to its essential role in enhancing urban safety and development. This interest can be attributed to the increased availability of geospatial and mobility data from a wide variety of sources, such as OpenStreetMap and knowledge graphs. Geospatial and mobility data enable the development of predictive models such as accident and crime prediction, enhancing urban safety, and planning. However, there are specific challenges to utilizing geospatial and mobility data when building predictive models. First, mobility data is typically sparse. Data sparsity occurs when spatio-temporal events, such as traffic accidents, are scarce and scattered across geographic regions. Due to data sparsity, predicting future events at specific locations becomes challenging. Second, geospatial and mobility data from multiple sources are often utilized by machine learning pipelines to generate latent representations. The latent representations derived from multimodal data are richer in context and beneficial for several predictive tasks. However, the diversity in data sources makes it challenging for machine learning pipelines to integrate these sources effectively, resulting in ineffective latent representations. Third, personal mobility data can contain sensitive information, such as visited locations, traveled routes, and driver profiles. Applications relying on personal mobility data require effective and robust methods to confirm provenance and authenticity. However, existing methods in the mobility domain are neither effective nor robust, which makes tracing personal mobility data challenging. This lack of traceability of personal mobility data limits its use in predictive model development. <br /> This cumulative thesis summarizes several novel methods to address these challenges. First, we propose a novel adaptive clustering method for accident prediction (ACAP) to address the challenge of data sparsity. ACAP aggregates traffic accident events dynamically with a grid-growing algorithm while considering underlying data distribution. Furthermore, ACAP enhances the prediction results of traffic accident events by focusing on adaptive task-specific regions. Second, to address the challenge of ineffective latent representations of geospatial regions, we propose a multimodal and multitask approach for region representation learning (MAGRE). MAGRE leverages multitask learning combined with attention-based fusion to enhance the effectiveness of region latent representations. These effective latent representations maintain the semantics for several downstream predictive tasks. Furthermore, the adaptive representations generated by MAGRE can be aggregated for user regions of interest of any shape and size without retraining. Third, to address the challenge of the lack of traceability of personal mobility data, we propose a novel watermarking approach for GPS trajectories called W-Trace. W-Trace embeds watermarks within GPS trajectories and is robust to adversarial modifications, enhancing traceability. In addition, W-trace maintains the utility of watermarked GPS trajectories for several downstream tasks. In summary, this thesis presents three novel contributions: i) an adaptive aggregation method for accident event data, ii) an effective and adaptive representation learning approach for geospatial regions, and iii) an effective, robust, and utility-preserving watermarking method for GPS trajectories.