Representation Learning Of Uncertainty On Multi-label Classification

The difficulty of multi-label classification lies in the multi-label nature of data,which characterizes the multiple correlations between instances and labels.Existing multi-label classification emphasizes the label correlation in logical label form.Either label distribution learning annotated by numerical label form or label enhancement learnt from logical label form remains uncertain to improve the multi-label classification performance.Many uncertainty factors,including noisy and the depending features,discriminative and missing labels,affect classification performance and challenge the data representation.A reasonable data representation is a foundation for effectively solving multi-label classification problems.Granular computing is a methodology that deals with the uncertainty of data analysis by hierarchical granulation.However,the detailed models can only characterize the concept uncertainty of single-label,which requires extensions before applying them to the multi-label case.This paper addresses classification uncertainty from both feature space and label space.It intensively studies the incompleteness and indiscernibility of labels.By leveraging the latent label correlations and analyzing distributions of misclassifications,we significantly improve the robustness and accuracy of the proposed four multi-label models.Concretely,our work has the following innovations:(1)The multi-label data from various sensors suffer from noise.We present a robust global and local label correlation(RGLC)model for multi-label classification.By leveraging subspace learning and latent label space learning,we decompose the feature space into clean and noisy parts and constrain self-reconstruction.A latent subspace preserves the inherent geometry of the data,and a compact and differentiated latent data representation avoid information loss.We consider a robust label correlation on the global and local view while reducing the degeneration of noisy features and missing labels.Experimental results demonstrate that the classifier learned in implicit space contributes to the classification performance.(2)The calibration threshold affects the final multi-label classification.We employ intuitionistic fuzzy sets and three-way decisions(IFTWLE)to examine the suitability of an arbitrary label to an instance and leverage the label enhancement for uncertain instances.An uncertainty measure quantifies the misclassification degree and improves the strength of label supervision from the logical label model.Concretely,the intuitionistic fuzzy set simultaneously quantifies the possibility of whether the label is associated with the instance or not.Then,we generate the components of instances with uncertainty at the instance level.However,the artificially annotated are often flawed.The IFTWLE uses machine learning instead of manual annotation to mine numerical labelling information and improve labelling supervision ability.Experimental results demonstrate that granular computing improves the accuracy as the instances with a high probability of prediction error are recognized and reclassified.(3)The pseudo-class distribution within the neighbourhood weighted by the separation margin provides insightful heuristic information in measuring uncertainty.In allusion to the multi-label nature,we devise an uncertain-prone measure to quantify the misclassification distribution across the label space.This measure supports a novel model called three-way decisions with label enhancement(3WDLE),where the model directly determines the label association via label-specific learning if the corresponding instances are with a lower value of the measure and takes a customized label enhancement procedure otherwise.Experimental results demonstrate that the accuracy is significantly improved as compared with learning with logical labels only.(4)Cost-sensitive learning has achieved impressive performance for classifications with imbalanced classes.The imbalanced class distribution and concept roughness in multi-label endow the applicability of decision-theoretic rough set,however,how to deal with the deferment remains unsolved.We develop a multi-threshold-based three-way label enhancement learning model(MGT-3WDLE)and take the label enhancement for those deferments.This means we can boost the overall classifications by replacing instances with roughness by refined supervision.Experimental results demonstrate that the simplified model generates more accurate classification results.