Track: Data Analytics and Big Data

Abstract

Incorporating data mining tasks in different levels of planning has become an essential tactic in business, industry, and other sectors. The rationales for implementing a data mining task, such as classification, significantly increase if the techniques used in classification provide optimal results. Logical Analysis of Data (LAD) is a classification approach known for its promising accuracy in classification and its capabilities in providing interpretable pattens. The main challenge in implementing LAD is the pattern generation problem. In this study, the pattern generation problem is solved to optimality to find maximum prime patterns. The proposed approach incorporates Benders decomposition and Apriori algorithm to generate prime patterns with high coverage from past observations. These patterns are then employed to build LAD classifiers that are used to assign class labels to unseen observations. Computational experiments conducted on seven public datasets show that results of LAD classifiers, established by using the proposed pattern generation algorithm, surpassed results of six machine learning algorithms implemented in IBM SPSS Modeler.