Performance Evaluation of n-Grams Ratio Features in Solving Multi-Classes Classification Problems

Abstract

We present experimental results that compare k-Nearest Neighbors (k-NN) and Support Vector Machines (SVM) algorithms to classifythe natural disasters multi-classes problem when n-grams ratio is used as the numerical features and compare three SVM approaches to classify the transportation accidents multi-classes problem when the same n-grams ratio is used as the features. In the former problem, we would like to investigate which of the two prominent algorithms have a better accuracy, while in thelatter problem, we would like to compare which of the three well-known SVM approaches for solving multi-classes problem performs best.In the natural disasters problem, the class labels are earthquakes, volcanic eruptions, flooding, landslides, and others, while in the transportation accidents problem, the categories are traffic collisions, maritime accidents, aviation accidents, and others. n-grams dictionaries of each category are used as the references in creating numerical features of the news articles. The results show that for the natural disasters problem, k-NN performs better than SVM and for the transportation accidents problem, DAGSVMoutperforms the other two SVM binary classification approaches.