Developing a Framework for Selecting an Appropriate Model based on the Ensemble Learning

We present a framework for selecting the optimal ensemble learning model based on ۱۴۳۳۱۰ crash observations with five classes. For non-ensemble models, we use five common models. ۲۶ ensemble learning models are derived from these five models. We suggest Diff۲ and Diff۳ measures for choosing the right model. The diff۲ is the difference between observations classified incorrectly as class ۱ and incorrectly classified as class ۳, ۴, or ۵. In Diff۳, we compare observations misclassified as class ۱ or ۲ with observations misclassified as class ۴ or ۵. We select the best model based on the following criteria: for class ۱, the largest R۱, for class ۲, the largest "Diff۲", for class ۳, a negative "Diff۳", and for classes ۴ and ۵, the highest "F۱-score". The paper ranks ۳۱ models based on its criteria. There are five ranking series. By comparing these rankings, we can determine, for example, whether the ۳rd best model for class ۱ corresponds to the best model for class ۲. For each model, ۵ "Ranks" are determined. Relationships between the ranks were then evaluated. Rank۱ and Rank۲, Rank۳ and ۵ have a relatively strong relationship. A negative and relatively strong correlation exists between Rankings ۲ and ۳, as well as Rankings ۲ and ۵.

Keywords:

Crash Severity Prediction , Machine learning model , Ensemble Voting Classifier , Imbalanced Multi-Class Classification