Fig 1.
The following three experimental processes are conducted: 1) data integration, 2) predictive model construction using Weka [37], and 3) result evaluation using sensitivity, specificity, and overall accuracy.
Table 1.
Independent and dependent variables collected from various sources and used in the proposed forecasting model.
Table 2.
Experimental procedure.
Fig 2.
The trends of four predictors, dengue cases, and the infection rates of female, male, and larvae mosquitoes, are plotted and compared.
This figure illustrates the trends of four parameters in Nakhon Pathom. The trends of these parameters in the other two provinces (Ratchaburi and Samut Sakhon) are similar but are not shown due to space limitations.
Table 3.
The effect of scaled and unscaled data on prediction performance and model construction time.
Fig 3.
The prediction performance comparison based on the MAE and forecasting accuracy.
Fig 3 (A) shows the MAE and the prediction accuracy with varying values of σ2 while C was fixed. In contrast, the value of C was varied and σ2 was fixed in Fig 3 (B) to determine the optimal values for the MAE and prediction accuracy.
Table 4.
Prediction performance comparisons of five techniques of two models.
Fig 4.
Prediction performance comparison based on specificity and sensitivity of the six models.
A higher specificity and sensitivity suggest a better prediction efficiency. Among the evaluated models, the SVM-R obtained the highest prediction performance by achieving sensitivity and specificity of 0.8747 using a 10-fold cross-validation performed on the test set data.
Fig 5.
Prediction performance comparison based on the maximum, average, and minimum accuracy of the six models compared with the accuracy of the training set.
The experiment was conducted 10 times using 10-fold cross-validation performed on the training set and the test set data using the SVM-L, SVM-P, SVM-R, NN, DT, and KNN techniques. All the collected results were averaged.