Autoencoder and Extreme Learning Machine Based Deep Multi-label Classifier

In this article, a novel deep neural network model has been proposed for multi-label (ML) classification. Deep architectures, well-known for their information extraction and learning capabilities, have been specifically considered here to deal with the complex nature of ML data. The proposed model broadly has two phases: feature extraction and cascaded ML classification. In the first phase, deep autoencoders (DAEs) have been employed to handle the large feature space of ML data. The subsequent phase of the network takes these reduced and enhanced features and passes them through a cascade of ML extreme learning machines (MLELMs) which intricately learns the input to output mapping and performs ML classification. This proposed stacked network is capable of handling a large feature space while performing fast classification. Experiments have been done with benchmark ML datasets by varying the size of the network components to determine the optimal depth of the proposed model. Analysis of DAE vs stacked autoencoder (SAE) has also been done for best feature extraction. Comparison with six state-of-the-art classifiers also shows the proposed model to have superior performance in most cases.