Information fusion using convolutional transform learning

Abstract

There are many real-world problems pertaining to the need for the fusion of information from multiple sources. Consider, for example, the problem of demand forecasting that requires estimating the power consumption at a future point given the available information till the current instant. At the building level forecasting, the inputs are usually power consumption, weather(temperature, humidity), and occupancy. This is a crucial problem in smart grids that ranges from planning electricity generation to preventing non-technical losses. Likewise, many such real-world examples can be cast as multi-channel information fusion based problems. Thus, we need the techniques whereby this varied nature of information from multiple sources can be combined/fused to predict some value(s) that can contribute significantly to future decision making. A bountiful of techniques have been proposed so far for multi-channel fusion, yet hardly any of them have been addressed as an end-to-end fusion formulation. Few of such solutions are based on techniques that include - Deep learning and Statistical Machine Learning (SML) algorithms. However, existing solutions related to deep learning paradigms involve Convolutional Neural Network (CNN). The latter might not guarantee distinct filters and hence, quality representations might not be obtained that could lead to redundancy. Secondly, CNNs are supervised and, therefore, require large labelled datasets that are not readily available in every other domain. Lastly, SML algorithms are largely prone to overfitting as these heavily rely on quality of features input. Thus, end-toend, multi-channel, both unsupervised and supervised Convolutional Transform Learning (CTL) based solutions are proposed that bridges all the gaps. The problems targeted lie under multiple domains including financial, biomedical and multiview image and text datasets. Firstly, this dissertation proposes unsupervised multi-channel fusion solutions to the problems in the financial domain - stock trading(trend prediction/classification) and stock forecasting(price prediction/regression) both of which include i time-series data. It preserves the true nature of time-series as univariate instead of frameworks treating them as 2D matrix/image. Also, the given solution is highly efficient in terms of training a single framework single framework and obtaining features that can be utilized for both classification and regression tasks. The latter benefit cannot be achieved with CNNs. Secondly, multiple information fusion problems are solved by giving supervised frameworks based on CTL and deep learning paradigms. Specifically, one of the frameworks is proposed to cater to the problem of stock trading that eliminates the issue of dead ReLU and guarantees representations that are more diverse helping in obtaining better performance over the state-of-art techniques. The latter has been validated via fair comparison with CNN where the proposed method supersedes it. Next, an information fusion solution is given that is supervised jointly trained and optimized approach based on CTL and Decision Forest (DF) for predicting Drug-Drug Interactions that could lead to Adverse Drug Reactions (ADRs) instead of utilizing them in a piecemeal fashion. Lastly, this thesis contributes to solve multiview clustering fusion problem handling the challenge of data-constrained scenarios. It involves the multiview datasets under image and text categories. A joint optimization of Deep CTL (DCTL) and K-Means clustering is proposed. It avoids the piecemeal approach and learns representations from clustering perspective with the help of K-Means clustering loss.