*5.2. Accuravy Test*

Tables 9 and 10 show the result of the 10-fold cross-validation of the proposed BN. The proposed BN produced 76.86% accuracy with the threshold value of 0.6. The specificity of the proposed BN (83%) was higher than the sensitivity (76.05%), which means that the proposed BN classifies better in the non-eating activity than the eating activity. Figure 5 shows the ROC (receiver operating characteristic) curve as the threshold for the eating probability decreases. The cost for decreasing the threshold was the smallest at the point 'threshold = 0.6', and where the threshold is lower than 0.2, the BN classified all activities as an eating activity. As shown in Figure 5, the AUC (area under curve) is fairly large, which supports the usefulness of the BN. Figure 6 shows the accuracy, sensitivity, and specificity of the various typical learning classifiers. We used the Weka 3.8.0 tool (of the university of

the Waikato, Hamilton, New Zealand) to analyze the results. Five classifiers have a large deviation between tests, as they tend to be overfitted to the train data; when the test data is composed mostly of similar data with the train data, their performance is very high, but in the other case, they are very low. The proposed BN, LR, and RF showed smaller deviations. The accuracy of the proposed BN was 7.54–14.4% higher than other classifiers. In the case of naïve Bayes and Adaboost, sensitivities are very high (96.15% and 95.91%, respectively), but specificities are also very low (37.68% and 53.77%, respectively), which means that the two classifiers classified most cases as an eating activity. For the multilayer perceptron (MLP), it showed good results among five other classifiers, but the time to build the model and classify was much higher than other methods. For the one-sample t-test, suppose the population has a normal distribution, and let the null hypothesis *H<sup>o</sup>* = 0*accuracy* < 0.80. With *X* = 0.7854,*s* = 0.386, t = −0.0378 > −2.262, and *H<sup>o</sup>* is rejected. When *Ho*0 =<sup>0</sup> *accuracy* > 0.90 , *t* = −0.2969 < −2.262 , so *Ho*0 is rejected and the proposed model is expected to have an accuracy of 0.8–0.9 for the population.

**Figure 5.** ROC curve for the proposed BN.

**Figure 6.** Ten-fold cross-validation for other typical classifiers (accuracy, sensitivity, specificity).


**Table 10.** Statistical indices of the results.

**Table 9.** Confusion matrix of the proposed BN.

