Semi − Disambiguated Partial Label Learning by Mixing Disambiguation and Non-disambiguation Strategies

Partial Label Learning (PLL) is employed to tackle the issue of ambiguous labeling, where each instance in the training set is associated with a set of labels, but only one label in the associated set is the ground-true label. Two common strategies, disambiguation and non-disambiguation, both achieve some improvements in tackling the PLL problem. However, they each have their own drawbacks. In this work, we develop a novel PLL method, namely, Semi-Disambiguated Partial Label Learning (SDPLL), to mix the two strategies with the goal of enhancing the strengths and making up for the weaknesses of both strategies. In specific, first we adapt label propagation to operate a coarse-level disambiguation, by which we eliminate low-confidence labels from the candidate label set to get the shrinking set of candidate labels. Next, we employ binary decomposition to conduct non-disambiguation treatment on the shrinking version of the dataset. Empirical experiments on both controlled UCI datasets and real-world datasets justify the superiority (at least the competitiveness) of SDPLL over several representative competing models.