Properties of the Decision Tree in the WEKA - Essay Example

?Examination of Each Attributes attribute “Diabetic” is taken as reference for all other attributes in the histogram. Apriori as the association algorithm along with decision tree J48 is used for the comparative analysis of the final_medicaldata file. Data mining package known as WEKA is used. Main characteristics are demonstrated in the following. J-48: J-48 as a decision tree is discussed. It depends upon the gain ratio in order to split up the attributes by using the depth-first strategy. Leaves are replaced by the sub trees through a pruning method that also reduced the over filtering. WEKA enables the one of two options such as pruned tree or not pruned tree as shown in the figure. Figure 1: Properties of the Decision tree in the WEKA (J48) In addition to above features, the WEKA also performs the test options for data use and data classification. Usage of the Training set: Evaluation of the classifier is based on the prediction of the instances of a class, which is trained on. Supplied Test: Evaluation of the classifier is also performed on the prediction of the instances of a class, which is loaded from the file. Cross Validation: By entering the number of fold into the text field of the Fold in the WEKA explorer the classifier is evaluated. Percentage Split: Data percentage is predicted by the evaluation of a classifier that takes the data out for the testing. The percentage field determines the specification of data held. During the training, data is used and provided the value of percentage field that makes the important part. Value of the reminder is reserved for the testing purposes. By the default, value of percentage split is stated as the 66%. Data about 34% is used for testing and remaining 66% is trained. Figure 2: WEKA with testing options Decision tree performance is determined by examining the cross validation and percentage split in the provided medical dataset. Usage of Cross Validation for generation of decision tree: In order to control the factors such as training’s set size and confidence by the process of cross validation, the flexibility is found in the decision tree of J48. Confidence factor is used to minimise or reduce the error rate of the classification. It is said that confidence factor is used to settle the problem of tree pruning. In order to classify the instances in a more accurate way, the classifier is given an opportunity by increasing the confidence factor and removing the noise of the training. The value of the confidence factor is 95% used for the dataset and leads to an outstanding outcome of 89.2% for the correct and classified instances and only 10.7% is the classified incorrectly as shown in the following figure. Figure 3: Use of cross validation based on the option J-48 decision tree to generate the results by WEKA. In the above figure, the calculation of J48 decision tree has been shown which includes correct values in details. Confusion Matrix is the important point in the given figure, which describes the ways in which a classifier makes an error in the prediction of a class type. According to Dunham (2003) the confusion matrix provides the correctness of the solution for the given classification problem. Another term used as an alternative to the confusion matrix is the contingency table. Two classes having a single dataset contain a column and two rows for the confusion matrix as shown in the figure 4. Predicted Actual Figure 4: Confusion Matrix Here FP represents the incorrectly classified number of negatives as positives and called as the commission errors. TP represents correctly classified number of positives. TN represents the correct classification of negative numbers, and FN shows the incorrect classification of positive numbers as negative. These are called as the omission errors. Predictive accuracy becomes the way for measuring the performance of a classifier. Predictive accuracy is known as the calculated success rate determined by the use of predictive accuracy as the confusion matrix. Predictive Accuracy = (TP + TN/TP+TN+FN+FP) *100 In figure 3 correct prediction of 323 attributes by the decision tree is shown, while it could not predict the 39 attributes of the class. As shown in the figure 3 of the confusion matrix the patients of class a with diabetes type1 only 26 attributes have been predicted correctly by the decision tree while 22 attributes predicted incorrectly. On the other hand, patients of diabetes type2 from class b only 297 attributes were predicted correctly by decision tree and 17 incorrectly predicted. Therefore, cross validation is used to measure the predictive accuracy with the help of J48. Figure 4: Measuring the predictive accuracy Figure 5: Decision Tree Use of Percentage Split to Generate the Decision Tree The percentage split technique utilizes the 34% data for testing and 66% for the data training as a default. There are only 200 instances used for the training and only 123 instances utilized for the testing. Currently at this stage, the J48 decision tree gives the better results than that of Cross Validation by the increase of 1% in the classification. Figure 6 shows that 111 instances are classified correctly out of the 123 with 90.2% test results. Only 12 instances are classified incorrectly. Looking upon the confusion Matrix in the figure 7, it is shown that patients of class a with diabetes type 1, the correct predicted attributes were 7 and 5 attributes were incorrect. While patients of class b with type 2, the 104 attributes were predicted correctly and 7 attributes incorrectly. Hence the calculation of the predictive accuracy is measured with the help of J48 through the Cross validation. Figure 6: Predictive Accuracy Figure 7: WEKA used for generation of results from Percentage Split option on the Decision tree of J-48. Association Rules As discussed on the association rules in the earlier chapters that used for the issuance of predictions for all attributes, it resulted into the establishment of relationship between attributes. Confidence and support are the key aspects to make the relationship among the different attributes. These key aspects reflected the advantages and assurance of the undiscovered rules. Apriori was used as the association rule that resulted into a good relationship between the attributes. Data was put into the ranges because the association rule did not work on the numerical data. Data was separated into ranges by the WEKA that also discretized the filter. Association rule was applied after the data was transformed into the ranges of bins. References Dunham, H.M. (2003). Data Mining Introductory and Advanced Topics, Pearson Education, Inc. Read More