Boosted lazy associative classifier

Abstract

Lazy machine learning algorithms have to learn every time it is been given a new example, however knowing which example is being classified gives them the advantage of adjusting their knowledge search accordingly. The Lazy Associative Classifier (LAC) is a rule-based demand-driven lazy machine learning algorithm that takes advantage of the information present in the example being classified by focusing its effort on inducing only rules that cover that particular example. Each rule comes from a frequent pattern present in the data. While, associative classifiers, in general, suffer from searching frequent patterns among the large number of existing patterns within the data, LAC breaks that problem down into many subproblems, solving one small problem at a time. Rule-based algorithms are often caught in the dilemma of not knowing the best way to combine their rules in order to form the best possible classifier. Usually, the choosing of a rule metric followed by a simple voting is used (as simple as assigning an importance -- or weight -- of one to each rule and averaging the accounts by each class). This approach is easily proven to be frail. Furthermore, LAC uses all rules available, which can be considered a large quantity of rules, regardless of their prediction quality. In this work we use a boosting algorithm known as Confidence-Rated Adaboost in conjunction with LAC to form a new, more accurate and smaller (in number of rules present in each model) classifier algorithm called BLACk. We prove that our approach is superior in terms of accuracy to LAC and other associative classifier. Nevertheless, we show that the built classifiers model are less complex compared to those built by LAC.