Exploring instance correlation for advanced active learning
- Publication Type:
- Issue Date:
Active learning (AL) aims to construct an accurate classifier with the minimum labeling cost by actively selecting a few number of most informative instances for labeling. AL traditionally relies on some instance-based utility measures to assess individual instances and label the ones with the maximum values for training. However, such approaches cannot produce good labeling subsets. Because instances exist some explicit / implicit relations between each other, instance-based utility measure evaluates instance informativeness independently without considering their interactions. Accordingly, this thesis explores instance correlation in AL and utilizes it to make AL’s more accurate and applicable. To be specific, our objective is to explore instance correlation from different views and utilize them for three different tasks, including (1) reduce redundancy for optimal subset selection, (2) reduce labeling cost with a nonexpert labeler and (3) discover class spaces for dynamic data. First of all, the thesis introduces existing works on active learning from an instance-correlation perspective. Then it summarizes their technical strengths / weaknesses, followed by runtime and label complexity analysis, discussion about emerging active learning applications and instance-selection challenges therein. Secondly, we propose three AL paradigms by integrating different instance correlations into three major issues of AL, respectively. 1) The first method is an optimal instance subset selection method (ALOSS), where an expert is employed to provide accurate class labels for the queried data. Due to instance-based utility measures assess individual instances and label the ones with the maximum values, this may result in the redundancy issue in the selected subset. To address this issue, ALOSS simultaneously considers the importance of individual instances and the disparity between instances for subset selection. 2) The second method introduces pairwise label homogeneity in AL setting, in which a non-expert labeler is only asked “whether a pair of instances belong to the same class”. We explore label homogeneity information by using a non-expert labeler, aiming to further reducing the labeling cost of AL. 3) The last active learning method also utilizes pairwise label homogeneity for active class discovery and exploration in dynamic data, where some new classes may rapidly emerge and evolve, thereby making the labeler incapable of labeling the instances due to limited knowledge. Accordingly, we utilize pairwise label homogeneity information to uncover the hidden class spaces and find new classes timely. Empirical studies show that the proposed methods significantly outperform the state-of-the-art AL methods.
Please use this identifier to cite or link to this item: