Event Data Mining and Classification from Multiple Streaming Sources

A novel solution to mining and classification of deformable events from multiple streaming image data sources is discussed. Observations of natural or manmade phenomenon using sensor networks or remote satellites are often acquired from various sensory measurement mechanisms placed at different locations. Furthermore, each source measures a different parameter (or different aspect of the phenomenon) resulting in strong and weak classifiers for different data sources. Previous solutions for multisource learning and mining are applicable to simultaneous co-registered data measurements that may not work in many practical applications. We discuss a new multisource classification solution using a generative model that reduces the multiple measurement spaces into a common feature space and maintains a unique feature space for each measurement source. A temporal classifier is used for temporal knowledge transfer by tracking the correspondence between consecutive measurements from different sources in the common feature space. In addition, an auxiliary source-specific classifier is used for each data source. A knowledge transfer solution based on a Bayesian approach is then used to fuse the transferred knowledge between the consecutive measurements from two sources (applied to the common feature spaces) with a source-specific classifier for the current observation (applied to the unique feature space) to ensure robust classification labeling even during instances when only measurements from a weak data source is used. Experimental results on a practical cyclone detection and tracking problem from multiple streaming remote satellite sources demonstrate the usefulness of our proposed approach.