Improvement of Pattern Quality for Logical Analysis of data(LAD) Method

Abstract

Logical analysis of data (LAD) is an important subfield of supervised machine learning and data mining. It is data analysis methodology, which use concepts from optimization, combinatorics and boolean functions. LAD is binary classification that used for boolean data with high explanatory power. Because patterns are the most important building blocks in LAD, they must carefully selected. One of the main drawbacks in LAD, which needs to be addressed, is the quality of the generated patterns and extract positive and negative patterns that can classified new observations with high accuracy. The main objective select some high quality patterns form the total number of patterns. For that purpose, the proposed methodology to address this specific issue is that study the LAD method and its refinements. Also, define a quality measures for pattern generation. Then, contribute to the improvement of the pattern selection procedures using optimization techniques such as Multiobjective Evolutionary Algorithms (MOEA), Pareto front and Mixed Integer-Linear Programs MILP in the General Algebraic Modelling System (GAMS) tools. Finally, use public benchmark datasets to validate results. RStudio used to develop new codes for generating all patterns and their characteristic and select an optimized subset of patterns using MILP (GAMS), MOEA and Pareto front to improve pattern quality. Then, classified test datasets to get very strong results with a high accuracy. Experiments on several binary datasets show the efficiency of used methods in order to reduce the number of generated patterns and increase the accuracy of classification model. In addition, feature selection algorithms applied on some datasets to select significant attributes and the result of these experiments are illustrated. Experiential study shows the high results in term of accuracy and pattern reduction in all proposed methods.