Multimodal temporal pattern mining

This paper proposes an approach for mining multimodal temporal patterns from multiple synchronous signal sequences generated by different modalities. The instances of the temporal patterns suffer from noise and non-linear temporal warping. There are non-pattern signal segments separating the instances of the temporal patterns in the whole signal sequences. Hidden Markov models with thresholds of supports are trained to capture the sub-patterns in each modality. The sub-patterns have overlaps and can be stitched together to form complete temporal patterns. The temporal information of the instances of the patterns in different modalities is then utilized to discover the multimodal temporal patterns.