4.1.1. Holdout and 10 Cross-Validations for Precision, Recall, F-measure, and ROC Evaluation

Table 3 presents the performance evaluation of END classifier with SVM and RF as base learner including precision, recall, F-measure and receiver operating characteristic (ROC) as the best classifier of the holdout method.

As shown in Table 3, the results of END classifier evaluation on holdout method obtained the best precision in activity A6 with 100% using SVM as a base learner compared to RF at 99.1%. For activity A5, both SVM and RF base learners achieved the precision results at 88.8%. But in activity A4, better precision was obtained by RF with 96.6% followed by SVM with 92.8%. For A3, A2 and A1, precision results show that SVM has given higher result range of 95.3% to 96.1% compared to RF with 89.6% to 92.7%. Recall results show that SVM obtained 99.10% and RF obtained 97.30% for activity A6. However, for activity A5, RF produced 94.6% compared to SVM with 91.5%. For A4, A3, A2, and A1, recall results show that SVM produced results ranging from 90.4% to 97.6% compared to RF with 90.1%

to 94.9%. Results of F-measure evaluation for activity A6 is 99.6% for SVM and 98.2% for RF. However, the F-measure results for activity A4 and A5 for RF are 93.2% and 91.6% respectively compared to SVM with 91.60% and 90.20%. Activities A1, A2, and A3 give SVM better F-measure results ranging from 94.7% to 96.9%, higher than RF range from 90.2% to 93.8%. RF gained greater results for ROC evaluation with results ranging from 99.4% to 100% compared to SVM which produced 95.2% to 99.8% for all activities. Table 4 presents the performance evaluation of Random subspace classifier with SVM and RF as base learner for 10-fold cross-validation method.


**Table 3.** Performance evaluation of each activity with random subspace classifier on the holdout method.

**Table 4.** Performance evaluation for each activity of a random subspace classifier on 10-fold cross-validation method.


As shown in Table 4, the results of Random subspace classifier using 10-fold cross-validation indicates SVM as base learner produces better precision range of results from 95.6% to 99% compared to RF with 90.3% to 98% for activities A1, A2, A3 and A6. However, RF has obtained better precision results of 94.2% and 95.9% compared to SVM with 93% and 93.5% for activities A4 and A5. For recall, activities A1, A2, A3, and A6 with SVM produced superior results from 94.7% to 99.4% compared to RF with 90% to 98.2%. However, RF has obtained better recall results of 94.2% and 95.9% compared to SVM with 93.9% and 92.5% for activities A4 and A5. The results of F-measure for activities A1, A2, A3 and A6 ranged from 95.1% to 99.2% for SVM and 91.2% to 98.1% for RF. The ROC results for RF were from 99.3% to 100% and 98.2% to 99.8% for SVM for all the activities. In dataset 1, overall results of 10-fold cross-validation model evaluation give better results compared to holdout for both base learners.
