Likelihood ratio-based probabilistic classifier

• Modern classification methods are likely to misclassify samples with rare but class-specific data that are more similar (less distant) to the data of another than the original class. This is because they tend to focus on the majority of data, leaving the information provided by the rare data practically ignored. Nevertheless, it is an invaluable source of information that should support classification of samples with such data, despite their low frequency. Current solutions considering the rarity information involve likelihood ratio models (LR). We intend to modify the existing LR models to establish the class membership for the analysed samples by comparing them with the samples of known class label. If two compared samples show similarities of rare but class-specific features it makes the analysed sample much more likely to be a member of this class than any other class, even when its features are less distant to the features of most samples from other classes. The fundamental advantage of the developed methodology is inclusion of information about rare, class-specific features, which is neglected by ordinary classifiers. Converting LR values into probabilities with which a sample belongs to the classes under consideration, generates a powerful tool within the concept of probabilistic classification.

Subject headings

NATURVETENSKAP  -- Data- och informationsvetenskap -- Datavetenskap (hsv//swe)

NATURAL SCIENCES  -- Computer and Information Sciences -- Computer Sciences (hsv//eng)

Keyword

Probabilistic classification; Likelihood ratio; Multimodal class distributions

Publication and Content Type

ref (subject category)

art (subject category)