Robust Learning from Multiple Information Sources

Abstract

In the big data era, the ability to handle high-volume, high-velocity and high-variety information assets has become a basic requirement for data analysts. Traditional learning models, which focus on medium size, single source data, often fail to achieve reliable performance if data come from multiple heterogeneous sources (views). As a result, robust multi-view data processing methods that are insensitive to corruptions and anomalies in the data set are needed.

This thesis develops robust learning methods for three problems that arise from real-world applications: robust training on a noisy training set, multi-view learning in the presence of between-view inconsistency and network topology inference using partially observed data. The central theme behind all these methods is the use of information-theoretic measures, including entropies and information divergences, as parsimonious representations of uncertainties in the data, as robust optimization surrogates that allows for efficient learning, and as flexible and reliable discrepancy measures for data fusion.

More specifically, the thesis makes the following contributions: 1. We propose a maximum entropy-based discriminative learning model that incorporates the minimal entropy (ME) set anomaly detection technique. The resulting probabilistic model can perform both nonparametric classification and anomaly detection simultaneously. An efficient algorithm is then introduced to estimate the posterior distribution of the model parameters while selecting anomalies in the training data.

2. We consider a multi-view classification problem on a statistical manifold where class labels are provided by probabilistic density functions (p.d.f.) and may not be consistent among different views due to the existence of noise corruption. A stochastic consensus-based multi-view learning model is proposed to fuse predictive information for multiple views together. By exploring the non-Euclidean structure of the statistical manifold, a joint consensus view is constructed that is robust to single-view noise corruption and between-view inconsistency.

3. We present a method for estimating the parameters (partial correlations) of a Gaussian graphical model that learns a sparse sub-network topology from partially observed relational data. This model is applicable to the situation where the partial correlations between pairs of variables on a measured sub-network (internal data) are to be estimated when only summary information about the partial correlations between variables outside of the sub-network (external data) are available. The proposed model is able to incorporate the dependence structure between latent variables from external sources and perform latent feature selection efficiently. From a multi-view learning perspective, it can be seen as a two-view learning system given asymmetric information flow from both the internal view and the external view.