Error-Robust Learning

Deviations from the clean assumptions (or assumed hypotheses) and/or presence of outlying data instances (or outliers), which are referred to as error, are very likely in data due to inevitable sensor failures or human factors. Error could exhibit in features and/or labels. Unfortunately, even a small amount of error can compromise the ability of existing machine learning algorithms to produce useful outcomes. Hence, it is extremely important to provide robustness to machine learning algorithms against error. The focus of this thesis is on error-robust learning aiming at producing useful outcomes when learning a task(s) even if the clean assumptions do not hold exactly and/or outliers are present. In this thesis, I will cover our solutions for the problems on the topic of error-robust learning. In the first work, I present a novel Markov chain method for error-robust multi-view clustering (EMVC) where static data is represented by multiple views/perspectives and may contain errors. In the second work, I present a novel method for outlier-robust multi-aspect streaming tensor completion and factorization (OR-MSTC) where streaming data may contain outliers and is prone to incompleteness. In the third work, I propose a novel method for outlier-robust multi-view subspace clustering with prior constraints (RMVSC). In the last work, I present a novel bi-level learning method for label-error robust multi-view semi-supervised classification (LR-MVSC) where a fraction of the data instances are incorrectly labeled. We evaluate the effectiveness and efficiency of the proposed methods on publicly available real-world as well as synthetic datasets.