A condensed hybrid decision tree for decision rules (CHDTDR)

A transparent system is much needed in fields like medical, business and banking because of its efficacy in comprehensive decision making, user convincing capability and manageability. It enhances accuracy, trust, accountability, informed decision-making, fraud prevention, regulatory compliance, user confidence, efficiency, security, accessibility, ethical standards, error reduction, and effective management, making it essential for effective management in such critical fields. Decision tree (DT) is defined as a data mining technique that produces transparent decision-making rules. The expression of the decision rules into a flow chart like representation makes the DT explicitly understandable and closely resemble human reasoning. However, sometimes DT generates some redundant and irrelevant decision rules which decrease the efficiency and comprehensibility. Therefore, this paper proposes a condensed DT named Condensed Hybrid Decision Tree for Decision Rules (CHDTDR) which is a transparent decision system with less number of rules, and overcomes the drawback caused due to irrelevant features. The irrelevant features are removed by a proposed wrapper feature selection (FS) technique named Stochastic Hill Climbing with K nearest neighbor (SHCK) which selects the best feature set and thereby significantly improves comprehensibility of the decision rules generated by the DT without compromising the accuracy. The novel feature selection approach uses a heuristic search method that uses a predefined maximum iteration for searches the best subset of features by removing the irrelevant features from the current feature set. The complete approach depends on the number of features and the predefined maximum iteration used giving a computation of product of both which varies depending on the size of the data. The proposed CHDTDR is compared with a simple DT as well as DTs trained after executing Sequential Forward Floating Search (SFFS) and Sequential Backward Floating Search (SBFS) feature selection methods in terms of classification accuracy, feature reduction, comprehensibility, precision and recall on 8 well known UCI datasets. Experimental results show that the proposed CHDTDR produces transparent DT with more efficient decision rules for decision making.