Exploring the combination of self and mutual teaching for tabular-data-related semi-supervised regression.

Semi-supervised learning (SSL) has drawn much attention since it can alleviate the predicament in which only limited labels can be accessed, with the help of numerous unlabeled data. Many deep neural networks (NNs) based SSL methods have been proposed recently, which mostly focus on image classification tasks, while fewer efforts have been taken to tabular data related regression scenarios. In this paper, to handle the semi-supervised regression problem for tabular data with NNs based methods, we empirically observe the demerits of previous methods and present a novel framework, to explore the combination of self teaching and mutual teaching. For one thing, self teaching is employed to each base learner by using consistency regularization so that the model is driven to be more stable and also robust to local perturbations. For another, the knowledge of other base learners is extracted and filtered to perform mutual teaching to avoid the confirmation bias problem and boost the training of each base learner. Here, we employ an uncertainty-based strategy for the filtering of knowledge in mutual teaching. Extensive experiments are performed on multiple real-world datasets to demonstrate the effectiveness of previous methods and the proposed framework, and further analyses are conducted to understand the influence of each factor in the proposed framework. • Data augmentation tricks may be unapplicable for tabular data. • The combination of self and mutual teaching for SSL regression tasks is explored. • Self teaching improves performance by enhancing robustness and stability. • Mutual teaching provides help by avoiding confirmation bias. • Experiments verify the effectiveness of the method. [ABSTRACT FROM AUTHOR]