CC-SLIQ: Performance Enhancement with 2k Split Points in SLIQ Decision Tree Algorithm

Decision trees have been found to be very effective for classification in the emerging field of data mining. This paper proposes a new method: CC-SLIQ (Cascading Clustering and Supervised Learning In Quest) to improve the performance of the SLIQ decision tree algorithm. The drawback of the SLIQ algorithm is that in order to decide which attribute is to be split at each node, a large number of Gini indices have to be computed for all attributes and for each successor pair for all records that have not been classified. SLIQ employs a presorting technique in the tree growth phase that strongly affects its ability to find the best split at a decision tree node. However, the proposed model eliminates the need to sort the data at every node of the decision tree; as an alternative the training data uses a k-means clustering data segmentation only once for every numeric attribute at the beginning of the tree growth phase. The CC-SLIQ algorithm inexpensively evaluates split points that are twice the cluster size k and results in a compact and accurate tree, scalable for large datasets as well as classified datasets with a large number of attributes, classes, and records. The classification accuracy of this technique has been compared to the existing SLIQ and Elegant decision tree methods on a large number of datasets from the UCI machine learning and Weather Underground repository. The experiments show that the proposed algorithm reduces the computation of split points by 95.62%, decision rules generated by 56.5% and also leads to better mean classification accuracy of 79.29%, thus making it a practical tool for data mining.