- Efficient time-series analysis can impact multiple application domains such as motif discovery in gene analysis or music data, extracting spectro-temporal patterns in acoustic scene analysis, or annotating and classifying electrical bio-signals (such as ECG, EEG, and EMG) for medical applications.
Time-series analysis involves a variety of tasks.
To predict future values of a time-series, many approaches focus on capturing the time dependence between samples, e.g., hidden Markov model or recurrent neural network. To learn a compact representation, transformation based approaches such as discrete Fourier transform (DFT), singular value decomposition (SVD) or convolutive neural network (CNN) are considered. To classify the time-series, one common approach is a distance-based K-nearest neighbor (KNN) with dynamic time warping (DTW) or edit distance. Other supervised approaches either use hand-crafted features or deep learning representation.
In time-series data, as in electrical bio-signals, acoustic scene, and music theme analysis, occurrence of patterns can be modeled as stationary or time-invariant. Discovering such patterns is the key to the analysis of time-series and can be further used for reconstruction or classification. Rich time-series data may contain multiple patterns associated with multiple labels (e.g., in acoustic scene analysis of an urban environment, vehicle sounds, bird sounds, and human speech may be observed in the time interval). To resolve the labels of individual instances, often an expensive fine-grain labeling process is required. To reduce the labor-intensive annotation efforts associated with labeling a signal at a time-instance level, coarse interval-level labeling is often considered.
In this context, we propose a framework that can efficiently model and analyze large-scale multi-channel time-series data and provide fine-grain label predictions from coarse interval-labeled data.
Since our focus is on time-series data with time-invariant events, we consider a convolutive modeling.
To extract time-invariant recurring patterns in time-series, we first propose a convolutive generative framework and use the resulting features for classification. To learn a time-instance label model in a weak-supervision setting efficiently, we propose novel dynamic programming approaches (using both a chain and a tree structure). Moreover, we extend the proposed weakly-supervised dictionary learning model for adapting both multiple clusters and multiple-scales. As future work, we present an application of the proposed approach to deep learning.