Handling obsolete information in classification: is there a one-size-fits-all strategy?

Add to your list(s) Download to your calendar using vCal

• Christoforos Anagnostopoulos, Imperial College London
• Friday 19 October 2012, 12:00-13:00
• FW26, Computer Laboratory.

If you have a question about this talk, please contact Ekaterina Kochmar.

Classifier performance is known to deteriorate over time, as a result of the training dataset becoming obsolete as time goes by. A particularly common instance of this is the phenomenon of “population drift”, whereby gradual changes in the population characteristics have a cumulative detrimental effect on the accuracy of the classifier over time. In online learning contexts, where the classifier is updated incrementally as new information arrives, this raises the question of whether novel information should complement, or altogether replace older data. In this talk, we illustrate how such decisions hinge not only on the precise nature of the evolution exhibited by the data, but also on the methodology that underlies the classifier: different classifiers are affected in different ways by drift. This observation has been studied by various authors in the streaming data literature, and in this talk we review and extend this work. We conclude by commenting on the suitability of one-size-fits-all strategies for handling drift, that monitor performance alone and are otherwise indifferent to the underlying methodology.

This talk is part of the NLIP Seminar Series series.

This talk is included in these lists:

• All Talks (aka the CURE list)
• Cambridge Centre for Data-Driven Discovery (C2D3)
• Cambridge Forum of Science and Humanities
• Cambridge Language Sciences
• Cambridge talks
• Chris Davis' list
• Department of Computer Science and Technology talks and seminars
• FW26, Computer Laboratory
• Graduate-Seminars
• Guy Emerson's list
• Interested Talks
• Language Sciences for Graduate Students
• NLIP Seminar Series
• PMRFPS's
• Simon Baker's List
• Trust & Technology Initiative - interesting events
• bld31
• ndk22's list
• ob366-ai4er
• rp587

Note that ex-directory lists are not shown.