Abstract
As one of the most promising classification schemes, association classification (AC) effectively integrates classification with association rule discovery, but it requires precisely labelled data which is usually expensive or hard to obtain in practice. This paper extends the AC to solve the classification of data with soft labels represented as belief functions defined over the set of classes. To characterize the imprecision of labels, a more general rule, called imprecise classification association rule (ICAR), is first introduced so that the consequent is more flexible, being any subset of the class set. Meanwhile, measures of support and confidence are proposed for ICAR by taking into account the belief functions encoded in soft labels. With the new rule structure, an association rule-based soft labelled classification method, called ARC-SL, is then developed to build a more accurate classification model containing three phases: entropy-based adaptive partition for deriving fuzzy regions of continuous attributes, Apriori-based rule mining for generating a set of ICARs with variable support and confidence thresholds, and rule pruning for discarding redundant or poor rules. Finally, a belief reasoning procedure is presented to classify each input instance through combining the activated rules in the framework of belief functions. Experiments on benchmark datasets and an application to facial expression recognition demonstrate the effectiveness of the proposed method.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have