Deep learning method for incomplete data classification

This paper proposes a novel deep learning technique called adaptive deep belief networks (ADBN) for classification with incomplete data, in which data values are partially observed. ADBN constructs the deep architecture using different kinds of restricted boltzmann machines (RBM) and refines the parameter space globally to maximize the separability. First, a novel adaptive RBM is used in the bottom layer of deep nets to address the incomplete input data. Secondly, we construct the hidden layers using a set of RBMs via greedy and layer-wise unsupervised learning. Thirdly, we construct the top layer using top RBM for classification. Finally, we fine-tune the parameter space of the deep architecture by global gradient-descent based supervised learning using a new loss function. ADBN not only inherits the powerful classification ability of deep belief networks (DBN), but also demonstrates the attractive characters of handling incomplete data. The empirical validation on the standard dataset shows that ADBN outperforms other incomplete data techniques in classification with different missing ratios and different missing types. Copyright © 2015 Binary Information Press.