USING GENETIC ALGORITHM FOR FEATURE SELECTION

Abstract

With the rise of machine learning, we are discovering better more complex, and performant models and new data-sets to use on those models. But as the dataset keep getting bigger and bigger, the time to train the model also increases. Feature selection helps in keeping in check this training time all the while improving the model performance by avoiding overfitting. As discussed in this thesis there are many types of feature selection algorithms. Each of those types has its pros and cons. We are focusing on genetic algorithm as the feature selection method of our choice. We compare genetic algorithm-based feature selection algorithm with five other feature selection methods. We found that the genetic algorithm based feature selection method has comparable performance while having lower error. We also use genetic algorithm-based feature selection with KNN to detect the mechanism of action of drugs and find that the Cross-validation score increases substantially from first generation to the twentieth generation. Showing that genetic algorithm as a feature selection method of choice is comparable if not better to other feature selection methods available.