Anytime Algorithms for Data Stream Classification with Application on Insect Classification

Abstract

Machine learning is one of the most influential research areas in Artificial Intelligence, with many practical applications in various domains. Initially, Machine Learning techniques assumed that the processing could be done in batch. The batch processing has the assumption that learning and classification can be performed without major restrictions in processing time. More recently, there has been a growing interest in application domains that generate data streams. Processing data streams has as main characteristic the need for answers that meet strict time constraints. For example, a classifier applied to a data stream must provide a response to a particular event before the next event occurs. Otherwise, some events of the stream may be left unclassified. Even more challenging is that many data streams generate events at a high variable rate of arrival, i.e. the time interval between two successive events can vary widely. An example of application that has the characteristics of a data stream with variable arrival time is the intelligent trap we are developing. This trap uses a sensor to identify and capture potentially harmful insect species for agriculture and public health. The classification of insects species requires algorithms able to provide answers under severe classification time constraints and with high variability in events arrival time. One possible solution to deal with these constraints is the use of anytime classifiers. These classifiers are able to provide responses with variable processing time; in turn they increase the response quality in function of processing time. In this project, we are interested in investigating anytime classification methods to handle data stream applied to the classification of insects. Our hypothesis is that anytime the versions of some traditional algorithms are able to provide efficient algorithms for classification of data stream without significant loss of accuracy. The methods investigated will be compared with traditional classifiers in terms of accuracy. Anytime algorithms will be compared with each other in terms of both classification efficacy and processing efficiency in real databases collected from the sensor, and benchmark databases.