A Framework for Learned Approximate Query Processing for Tabular Data with Trajectory

Approximate query processing has been well established for enhancing performance of aggregation queries on ever-increasing big data by statistically equivalent approximations. Recent popularity of mobile devices creates tremendous spatio-temporal data that require different treatment than relational ones. Among spatio-temporal data, we focus on trajectories in a tabular form and analyzes the problem, its requirements, and suggest a general-purpose framework for learned approximate query processing by providing a common encoding/embedding layer for embracing diverse state-of-the-art ML models, on top of which resides a probabilistic circuit for efficiency and efficacy with error bounds.