Research Project: Human Behavior Modeling via Mobile Sensing
- Proposed the first computational framework for modeling rhythms of human behaviors from multimodal sensor signals, including three main components: mobile sensor data processing, rhythm discovery & modeling, and machine learning prediction of health and wellbeing outcomes. The framework achieved the best performance in predicting depression and productivity.
- Applied the non-parametric ANOVA and recurrent neural network (RNN) change point detection algorithms for sequential sensor data to identify unhealthy human behavioral events.
- Proposed a novel Transformer model called TrFHB for the cyclic time series prediction and achieved better performance than baseline deep learning models (e.g., LSTM and DeepAR).
- Proposed a parameter-free, unsupervised human behavior clustering method called Wavelet Transfer Learning (WTL). The method creates image-based representations of multimodal sensor data streams that allow for extracting granular categorization of the behavior that could not be revealed in existing approaches.
