Machine learning models for hierarchical multilabel classification (HMC) typically achieve low accuracy. This is because such models need not only predict multiple labels for each data instance, but also ensure that predicted labels conform to a given hierarchical structure. Existing state-of the-art strategies for HMC decouple the learning process from ensuring that predicted labels reside in a path of the hierarchy, thus inevitably degrading the overall classification accuracy. To address this limitation, we propose a novel loss function, which enables a model to encode both a global perspective of the class hierarchy, as well local class-relationships in adjacent hierarchical levels, to ensure that predictions align with the class hierarchy, both during training and testing. We demonstrate the superiority of the proposed approach against multiple state–of–the–art methods for HMC on 20 real-world datasets.

Existing approaches for multi–label classification are trained offline, missing the opportunity to adapt to new data instances as they become available. To address this gap, an online multi–label classification method was proposed recently, to learn from data instances sequentially. In this work, we focus on multi–label classification tasks, in which the labels are organized in a hierarchy. We formulate online hierarchical multi–labeled classification as an online optimization task that jointly learns individual label predictors and a label threshold, and propose a novel hierarchy constraint to penalize predictions that are inconsistent with the label hierarchy structure. Experimental results on three benchmark datasets show that the proposed approach outperforms online multi–label classification methods, and achieves comparable to, or even better performance than offline hierarchical classification frameworks with respect to hierarchical evaluation metrics.

Data quality is of paramount importance in the training process of any machine learning model. Recently proposed methods for noisy learning focus on detecting noisy labeled data instances by using a fixed loss value threshold, and exclude detected noisy data instances in subsequent training steps. However, a predefined, fixed loss value threshold may not always be optimal, and excluding the detected noisy data instances can hurt the size of the training set. In this paper, we propose a new method, NDCC, that automatically selects a loss threshold to identify noisy labeled data instance, and uses counterfactual learning to repair them. To the best of our knowledge, NDCC is the first work to explore the feasibility of using counterfactual learning in the noisy learning domain. We demonstrate the performance of NDCC on Fashion–MNIST and CIFAR–10 datasets under a variety of label noise environments. Experimental results show the superiority of the proposed method compared to the state–of–the–art, especially in the presence of severe label noise.

Most supervised classification methods assume perfect training data, although this is not usually the case in the real–world. Meanwhile, counterfactual data generation approaches have emerged as a way to provide post–hoc explanation of decisions made by classification models. However, such approaches highly rely on the classification model output since different outputs lead to alternative, or even contradicting explanations. This work proposes a plug–and–play framework to learn a robust classification model in the presence of noisy labeled data and provide actionable suggestions for undesirable decisions (e.g., loan application rejection) made by a given classification model. The framework’s generalizability is demonstrated by considering alternative noisy label detection and counterfactual explanation methods, as well as diverse supervised classification models. The framework’s superiority against several baselines is demonstrated using three benchmark datasets.

One of the most challenging machine learning problems is a particular case of classification in which classes are hierarchically structured and data instances can be assigned multiple labels residing in a path of the hierarchy. In this paper, we propose hierarchy–aware multiclass AdaBoost, allowing for the first time weak classifiers in an ensemble learning setting to be trained for hierarchical multiclass classification while incorporating a hierarchy–aware loss function directly into the training process. Experimental results on numerous real–world datasets show that, despite its simplicity, the proposed algorithm outperforms all baselines, arising as the state of the art in hierarchical multiclass classification.

Although data quality is of paramount importance in algorithmic decision–making, most existing methods for supervised classification use training data without ever questioning their fidelity. At the same time, counterfactual explanation approaches widely used for post–hoc explanation of algorithmic decisions may result in unrealistic recommendations when left unconstrained. This work highlights a significant research problem, and introduces a novel framework to improve supervised classification in the presence of untrustworthy data, while offering actionable suggestions when an undesirable decision has been made (e.g., loan application rejection). Evaluation results spanning datasets from different domains demonstrate the superiority of the proposed approach, and its comparative advantage as the percentage of mislabeled instances increases.