Sparse Maximum Margin Learning from Multimodal Human Behavioral Patterns

We propose a multimodal data fusion framework to systematically analyze human behavioral from specialized domains that are inherently dynamic, sparse, and heterogeneous. develop two-tier architecture of probabilistic mixtures, where the lower tier leverages parametric distributions exponential family extract significant patterns each modality. These then organized into dynamic latent state space at higher fuse different modalities. In addition, our jointly performs pattern discovery maximum-margin learning for downstream classification tasks by using group-wise sparse prior regularizes coefficients classifier. Therefore, discovered highly interpretable discriminative support tasks. Experiments on real-world medical psychological demonstrate discovers meaningful with improved interpretability prediction performance.