Classification Rule Mining with Iterated Greedy

Abstract

Abstract. In the context of data mining, classification rule discovering is the task of designing accurate rule based systems that model the useful knowledge that differentiate some data classes from others, and is present in large data sets. Iterated greedy search is a powerful metaheuristic, successfully applied to different optimisation problems, which to our knowledge, has not previously been used for classification rule mining. In this work, we analyse the convenience of using iterated greedy algorithms for the design of rule classification systems. We present and study different alternatives and compare the results with state-of-the-art methodologies from the literature. The results show that iterated greedy search may generate accurate rule classification systems with acceptable interpretability levels.

Multi-view Genetic Programming Learning to Obtain Interpretable Rule-Based Classifiers for Semi-supervised Contexts. Lessons Learnt

“…We think that this line of research is worthy of further studies, particularly given that the combination of multi-view learning and rule-based classifiers has not been very much studied. We intend to explore the following avenues: (1) to consider other diversity measures and application schemes that might favor the exploitation of the complementary principle; (2) to search for rule-based classifiers under the multi-view methodology and other search models apart from genetic programming, such as iterated greedy [21]; and (3) to analyze whether better rule-based classifiers can be obtained under the multi-view paradigm where other views make use of other and more powerful base classifiers, such as neural networks.…”

“…Comprehensibility is one of the benefits that rule-based systems possess and have attracted the attention of researchers [17][18][19]21,48]. One of the most common ways to evaluate the interpretability of rule-based systems is counting the number of conditions present in their rules.…”

“…In the context of supervised learning, the IF part, or antecedent, defines attribute conditions that patterns must satisfy in order to be covered by the rule; whereas the THEN part, or consequent, predicts the class of these covered patterns. One of the major benefits of rule-based systems, in contrast to other machine learning tools, is their comprehensibility, because rules are easily interpretable for experts in the addressed problems [17][18][19][20][21]. Notice here that one particularity of rule-based systems is that they define their own set of attributes whose decisions are based on.…”