Combined Firefly Algorithm-Random Forest to Classify Autistic Spectrum Disorders

Abstract

Early diagnosis of autistic spectrum disorder, an imperfect neurological development condition, is one way to reduce the sufferer condition. However, the diagnosis of ASD is costly. A popular classification model based on a machine learning technique, such as random forest, can reduce the cost. In general, an RF that is designed by a domain expert gives high accuracy for various datasets. Unfortunately, RF commonly produces a low F1-score for an imbalanced-class dataset. Therefore, in this paper, a firefly algorithm, one of the popular swarm intelligence algorithms, is exploited to automatically design an optimum RF. First, a decision tree is formed based on random features chosen by RF. The decision trees have different features, which cause RF to have new knowledge to classify data continually. The feature used to form a decision tree is 20% of the total attributes. This decision tree is then formed into a forest. Finally, it classifies data using a voting scheme. In FA-based optimization, an individual firefly represents one decision tree. The objective function of a firefly is based on its accuracy. An evaluation using the ASD datasets shows that the proposed combination of FA and RF (FARF) performs better than the original RF for a decision tree of 30. FARF reaches an accuracy of 94.32% and F1-scores of 35.67%, while RF gives an accuracy of 90.78% and F1-scores of 34.09%.