Advances in Temporal Data Mining for Clustering, Classification, and Prediction Purposes: Integrating Domain Knowledge within Temporal Learning

There is a growing need for classification, prediction, andclustering of large numbers of multivariate time-series data: Integrating information from multipleIntelligence sources, predicting an outcome given a series of measurements, or noticing that thedata represent several different temporal trajectories. However, most data mining techniques treattemporal data either as an unordered collection of events, ignoring its temporal information, or asa simple sequence of symbols, ignoring the time dimension~s properties. Despite advances in thisarea, nearly all proposed algorithms in sequential pattern mining focus on time-point-based data, in whichevents occur at single time instants, and typically focus on raw data rather than on higher level concepts.However, events are not necessarily instantaneous. They can occur a time interval (e.g., the system was downfor 15 minutes). Current algorithms, including Deep Learning networks, often assume that the data in thedifferent information channel are sampled at a constant frequency, and are synchronized. Most algorithmsalso focus on fully automated learning rather than on integrating existing domain knowledge.