The vast amounts of spatio-temporal data generated by a variety of devices are best utilized in conjunction with relational, tabular data rather than being referenced separately. To enhance the efficiency of data analysis, learned models are frequently used to approximate query results by increasing responsiveness at the cost of some accuracy. Machine learning techniques exist for spatio-temporal data and tabular data separately, but it is not straightforward to represent the combined data in a unified learned model. This paper explores the challenge of learning from heterogeneous data, particularly trajectory and tabular data, and proposes approaches for representation learning using probabilistic circuits and deep neural networks.

Huge amounts of data are ceaselessly being generated by a variety of devices, and the processing efforts for their collection and analysis grows exponentially as well. Storing them in one place and getting exact answers is almost impractical. Furthermore, computing aggregation and statistics that most exploratory data analysis would require imposes a heavy burden on networking and computing infrastructures. By adopting the edge/fog computing paradigm that has recently been developing can reduce such overheads by offloading jobs from central clouds to edge devices. We try to go one step further in this direction by approximating aggregate values and statistics for data analysis using tractable probabilistic models and optimizing network performance. This paper evaluates our preliminary result of our on-going project that was gained by fast-prototyping using Sum-Product Networks.