A Reduced KELM model using DBSCAN Clustering algorithm for centroid selection

• Published in: 2018 5th International Conference on Signal Processing and Integrated Networks (SPIN) pp. 702 - 707
• Main Authors: Jain, Sukirty, Shukla, Sanyam
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.02.2018
• Subjects: Clustering algorithms; Computational modeling; Data mining; DBSCAN clustering algorithm; Extreme Learning Machine; Kernel; Kernel Extreme Learning Machine; Kernel matrix; Neurons; Testing; Training
• DOI: 10.1109/SPIN.2018.8474165

Extreme Learning Machine (ELM) is a single layer feedforward neural network (SLFN) and a popular classifier for classification and regression problems. It is unstable due to random initialization of weights between the hidden layer and the input layer. To overcome this problem of instability, kernelized ELM with kernels has been developed. Gaussian kernel Extreme Learning Machine (KELM) is one of the stable but computationally complex variant of ELM in which the number of hidden layer neurons is equal to the number of input instances. Therefore, to reduce this computational complexity, a new variant of KELM i.e. reudced KELM has been proposed in the literature. It randomly selects the number of centroids from the training data set such that the number of centroids is always less than the number of input instances. This work develops a new reduced version of KELM using density-based clustering algorithm (DBSCAN). DBSCAN is a non-spherical clustering algorithm which is used to identify the number of concepts. Each concept is represented using the centroid of the cluster. In proposed KELM model, the number of hidden layer neurons is equal to the number of clusters formed by the DBSCAN clustering algorithm. Experiments have been performed on 16 datasets drawn from the KEEL data set repository. The result shows that the developed reduced KELM model gives better performance as compared to the traditional KELM in terms of reduced training time and testing time.