Learning with Kernels for Streams of Structured Data

Abstract: Learning from streaming data represents an important and challenging task. Maintaining an accurate model, while the stream goes by, requires a smart way for tracking data changes through time. This challenge is harder if the stream is constituted by structured data, such as trees and graphs.
In the talk, I motivate why it makes sense to consider streams of structured data. Moreover, after recalling the basic computational requirements for processing streams, I discuss the additional difficulties that are encountered when using kernels for structured data.
Subsequently, I report on recent advances on the efficient use of kernels for structured data on streams of trees and graphs. All these techniques are based on efficient representation of the learning model and/or of the feature space.
Finally, I close the talk with an example of a promising direction of research where a well known data mining technique, i.e. Lossy Counting, can be integrated into the learning process so to dramatically speed up training while preserving performances.