Semi-Supervised Learning of a Classifier from a Statistical Perspective
Date: 23 July 2021, Friday
Time: 4pm AEDT
Speaker: Prof Geoffrey McLachlan (University fo Queensland)
Abstract:
With the considerable interest on machine learning these days, there is increasing attention being given to a semi-supervised learning (SSL) approach to constructing a classifier. From a statistical perspective, it goes back over 50 years (McLachlan, 1975, JASA). As is well known, the (Fisher) information in an unclassified feature with unknown class label is less (considerably less for weakly separated classes) than that of a classified feature which has known class label. Hence in the case where the absence of class labels does not depend on the data, the expected error rate of a classifier formed from the classified and unclassified features in a partially classified sample can be relatively much greater than that if the sample were completely classified (Ganesalingam and McLachlan, 1978, Biometrika). On treating the labels of the unclassified features as missing data and adopting a framework for their missingness, it is shown that the performance of the Bayes’ classifier can be improved to an extent where the SSL rule so produced can outperform the rule based on the sample if it were completely classified (Ahfock and McLachlan, 2020, Statist Comput). This is a most surprising result. It can occur in situations where the unclassified features tend to fall in overlapping regions of the classes in the feature space; that is, for features that are difficult to classify. Such features tend to have relatively high entropy and so it is proposed that the probability a class label is missing be modelled as a function of the entropy of the associated feature vector. This is joint work with Daniel Ahfock.
Link: https://au.bbcollab.com/guest/fcf219c74ac743e89565a9e6e8d349a9
Date: 23 July 2021, Friday
Time: 4pm AEDT
Speaker: Prof Geoffrey McLachlan (University fo Queensland)
Abstract:
With the considerable interest on machine learning these days, there is increasing attention being given to a semi-supervised learning (SSL) approach to constructing a classifier. From a statistical perspective, it goes back over 50 years (McLachlan, 1975, JASA). As is well known, the (Fisher) information in an unclassified feature with unknown class label is less (considerably less for weakly separated classes) than that of a classified feature which has known class label. Hence in the case where the absence of class labels does not depend on the data, the expected error rate of a classifier formed from the classified and unclassified features in a partially classified sample can be relatively much greater than that if the sample were completely classified (Ganesalingam and McLachlan, 1978, Biometrika). On treating the labels of the unclassified features as missing data and adopting a framework for their missingness, it is shown that the performance of the Bayes’ classifier can be improved to an extent where the SSL rule so produced can outperform the rule based on the sample if it were completely classified (Ahfock and McLachlan, 2020, Statist Comput). This is a most surprising result. It can occur in situations where the unclassified features tend to fall in overlapping regions of the classes in the feature space; that is, for features that are difficult to classify. Such features tend to have relatively high entropy and so it is proposed that the probability a class label is missing be modelled as a function of the entropy of the associated feature vector. This is joint work with Daniel Ahfock.
Link: https://au.bbcollab.com/guest/fcf219c74ac743e89565a9e6e8d349a9
Video: