On the other hand, quality of classification is the percentage of correctly classified cases. supplier Maraviroc In this study, 91.9% of cases are correctly classified, indicating well-performed robustness of the rough sets model. Table 4 Approximation results. The reducts from the training
set are calculated using the computationally efficient genetic algorithm option in Rosetta. The genetic algorithm is a heuristic for function optimization and promotes “survival of fittest” . In total more than 3000 reducts are calculated. The length of the reducts is 2~12 attributes. It represents that any attribute is necessary for perfect approximation of the decision classes and removal of any of them leads to the decrease of the quality of approximation. 5.2. Decision Rule Induction Based on the concepts of indiscernibility relations, set approximation,
and attribute reduction, the training set is analyzed and over 40,000 rules are generated. This means that most rules are supported by just one or two objects. In fact, the highest support for an exact rule in this data is only 64 objects. The top five supported rules are shown in Table 5. Table 5 Top supported induced decision rules. 5.3. Validation Confusion (or misclassification) matrix measures the effectiveness of the mode choice modeling. Table 6 presents confusion matrix induced by the model for the testing set. In a confusion matrix, the sum on each row or column represents the actual or predicted number of observations for each mode. The main diagonal cells give the match number between reality and prediction and off-diagonal provides the erroneous classification. The accuracy and coverage for each mode appear in the table as the index of prediction performance. Table 6 Confusion matrix generated by rough sets model. Overall, the rough sets model has a good accuracy prediction, with
the overall accuracy (hit ratio) up to 77.3%. The misclassification results reflect that it cannot distinguish between the SOV and car modes well in the fact that many observations under these two modes are mutually misclassified. This phenomenon indicates that the SOV and car modes, which share Entinostat household, individual and travel attributes, exhibit more homogeneity within the explanatory variables than other modes. The model yields the highest prediction accuracy for foot with the rate up to 91.4%, showing most of the observations choosing the foot mode are not misclassified as other modes. However, the bicycle is underestimated heavily. A large part of the misclassified observations of the bicycle mode goes to the SOV mode, which may imply some unobserved similar preferences between SOV travelers and bicycle users. On the other hand, the rough sets model made acceptable predictions of the mode choice distribution on the coverage level.