Modified iterative dichotomiser 3 (ID3) algorithm applied in diabetes risk detection

ABSTRACT: This study presents a modified Iterative Dichotomiser 3 (ID3) decision tree algorithm, designed to address multi-value bias, equally important attribute problem, and overfitting where these challenges impact the tree’s classification accuracy. This aim to enhance the algorithm by improving its attribute selection, handling mechanism for tied attributes, and applying a regularization technique for better generalization. The modified ID3 utilized mutual information-based information gain for attribute selection, incorporated purity calculation for tie-breaking situations, and introduced the concept of dropout regularization to mitigate overfitting. Testing was conducted on a diabetes dataset containing 520 instances with 16 features of categorical values. The model was evaluated using standard performance measures (accuracy, precision, recall, F1 score) and compared to the traditional ID3 and other modified versions. The modified algorithm achieved an average accuracy of 97% which surpassed the traditional ID3 algorithm and other modified ID3 algorithms. These findings reveal an average accuracy improvement of approximately 1% to 3% across training holdouts of 50% to 90% and various dropout rates of 0.1 to 0.4. Additionally, the modified ID3 algorithm produced less average number of nodes compared to the traditional ID3 algorithm, ranging from a difference of 6 to 22 nodes. Overall, the modifications that were implemented further improved the traditional ID3 algorithm’s classification performance, producing a more accurate and reliable decision tree.

Filipiniana