Semi-supervised learning of local structured output predictors

Abstract

In this paper, we study the problem of semi-supervised structured output prediction, which aims to learn predictors for structured outputs, such as sequences, tree nodes, vectors, etc., from a set of data points of both inputoutput pairs and single inputs without outputs. The traditional methods to solve this problem usually learns one single predictor for all the data points, and ignores the variety of the different data points. Different parts of the data set may have different local distributions, and requires different optimal local predictors. To overcome this disadvantage of existing methods, we propose to learn different local predictors for neighborhoods of different data points, and the missing structured outputs simultaneously. In the neighborhood of each data point, we proposed to learn a linear predictor by minimizing both the complexity of the predictor and the upper bound of the structured prediction loss. The minimization is conducted by gradient descent algorithms. Experiments over four benchmark data sets, including DDSM mammography medical images, SUN natural image data set, Cora research paper data set, and Spanish news wire article sentence data set, show the advantages of the proposed method.

Semi-supervised trees for multi-target regression

“…Unfortunately, this has not been the case for SOP tasks, although the need for SSL is even stronger there: The labeling process is even more expensive and laborious in SOP, since several simple/primitive labels (or one structured/complicated label) need to be provided for each example. Furthermore, the existing SSL methods for SOP largely deal with discrete types of outputs, such as a recent method proposed by Du [20]. There are only a few examples of SSL methods dealing with the task of MTR [35,23,31,10].…”