ACKEM: Automatic Classification, Using KNN Based Ensemble Modeling

Over the last couple of years, Deep Learning (DL) methods for objects and features classification have been shown to overcome previous state-of-the-art classification techniques in multiple areas, such as image classification and speech recognition. In our previous paper MESRS – Model Ensemble Speech Recognition System, we have described a unique speech recognition system for automatic classification of voice commands. The work described in this paper, presents a novel method for classification that continues our previous work by extending the system-supported input to the image space, not just the audio space. Aside from supporting multiple input types, this paper also describes an automated method of models ensemble based on the K-Nearest Neighbors algorithm. The automatic method of ensemble selection was added in order to improve the system’s running times and achieve the highest possible accuracy results. The work in this paper shows that applying dynamic input-based classification over multiple architectures can significantly improve the final classification results. Since different models with different architectures could achieve different results on different inputs, the task of producing the best results could be achieved by selecting the best fitted model for the given input. This method was tested over multiple datasets including Chest X Ray Pneumonia Dataset, Malaria Cells Dataset, Road Potholes Dataset, and the Voice Commands Dataset which also served us in our previous work. This paper proves that our method works and has the ability to improve the classification quality on top of all of the above datasets. Our results were compared with previous results obtained by similar works on top of the above datasets and a significant improvement was shown for all of the tested datasets. These findings prove the effectiveness of our method and motivate us to develop it further, in order to achieve even better results in future work.