A Hybrid Generative Model based on Diffusion and Graphs for Cross-Correlated Synthetic Multivariate Time Series Data

The prediction of missing sensor data in human activity recognition is an active field of research that is being targeted with generative models for synthetic data generation. In contrast to most previous approaches, which focus on extrapolation or prediction of data samples of a particular sensor, we target the generation of data for new sensor locations or modalities, i.e., different body locations or sensor modalities for which data had not been recorded in the first place. This is possible due to inherent correlations in the motion of human body parts. Particularly, from a larger body of training data, comprising diverse kinds of sensor body locations and modalities, we aim to learn correlations between body parts and sensor modalities, which are used to train a generative model to predict from existing sensor data of an individual subject, sensor modalities at different body locations. We also evaluate existing approaches proposed in the literature for their suitability in this scenario. This paper proposes a hybrid machine learning model based on diffusion and graph neural networks.