On-line Linear Discriminant Classifier

We present an on-line Linear Discriminant Classifier for streaming data (O-LDC). This is an adaptation of the Linear Discriminant Classifier, with the class means and the inverse covariance matrix re-calculated after each new data point. The classifier satisfies the properties of an on-line classifier; it learns from a single pass through the data, uses limited memory and processing power, and exhibits any-time learning. We compare the O-LDC with on-line versions of the Perceptron and balanced Winnow classifiers. Comparisons are carried out across a series of static data sets made up of two classes. The O-LDC shows higher accuracy and a better learning rate than its counterparts. As a second task we consider delayed labelling. We propose two strategies. The passive strategy waits for the correct label of a data point before using it to update the classifier. The aggressive strategy, makes use of naïve labelling, using the predicted label of a data point to update the classifier. The strategies are compared across a series of static data sets. The final accuracy of both strategies was comparable, though the passive strategy showed a better learning pattern.

Autorentext

Catrin Plumpton has an MMath degree and an M.Sc. in Computer Science from the University of Wales, Bangor. Having completed her M.Sc. in 2007 she then began a Ph.D. at Bangor University (the university having changed its name upon becoming autonomous in the same year).