Human activity recognition (HAR) using wearable sensors has advanced through various machine learning paradigms, each with inherent trade-offs between performance and labeling requirements. While fully supervised techniques achieve high accuracy, they demand extensive labeled datasets that are costly to obtain. Conversely, unsupervised methods eliminate labeling needs but often deliver suboptimal performance. This paper presents a comprehensive investigation across the supervision spectrum for wearable-based HAR, with particular focus on novel approaches that minimize labeling requirements while maintaining competitive accuracy. We develop and empirically compare: (1) traditional fully supervised learning, (2) basic unsupervised learning, (3) a weakly supervised learning approach with constraints, (4) a multi-task learning approach with knowledge sharing, (5) a self-supervised approach based on domain expertise, and (6) a novel weakly self-supervised learning framework that leverages domain knowledge and minimal labeled data. Experiments across benchmark datasets demonstrate that: (i) our weakly supervised methods achieve performance comparable to fully supervised approaches while significantly reducing supervision requirements; (ii) the proposed multi-task framework enhances performance through knowledge sharing between related tasks; (iii) our weakly self-supervised approach demonstrates remarkable efficiency with just 10% of labeled data. These results not only highlight the complementary strengths of different learning paradigms, offering insights into tailoring HAR solutions based on the availability of labeled data, but also establish that our novel weakly self-supervised framework offers a promising solution for practical HAR applications where labeled data are limited.

Machine learning-based decision support systems bring relief and support to the decision-maker in many domains such as loan application acceptance, dating, hiring, granting parole, insurance coverage, and medical diagnoses. These support systems facilitate processing tremendous amounts of data to decipher the patterns embedded in them. However, these decisions can also absorb and amplify bias embedded in the data. To address this, the work presented in this paper introduces a new fairness measure as well as an enhanced, feature-rich representation derived from the temporal aspects in the data set that permits the selection of the lowest bias model among the set of models learned on various versions of the augmented feature set. Specifically, our approach uses neural networks to forecast recidivism from many unique feature-rich models created from the same raw offender dataset. We create multiple records from one summarizing criminal record per offender in the raw dataset. This is achieved by grouping each set of arrest to release information into a unique record. We use offenders’ criminal history, substance abuse, and treatments taken during imprisonment in different numbers of past arrests to enrich the input feature vectors for the prediction models generated. We propose a fairness measure called Bias Parity (BP) score to measure quantifiable decrease in bias in the prediction models. BP score leverages an existing intuition of bias awareness and summarizes it in a single measure. We demonstrate how BP score can be used to quantify bias for a variety of statistical quantities and how to associate disparate impact with this measure. By using our feature enrichment approach we could increase the accuracy of predicting recidivism for the same dataset from 77.8% in another study to 89.2% in the current study while achieving an improved BP score computed for average accuracy of 99.4, where a value of 100 means no bias for the two subpopulation groups compared. Moreover, an analysis of the accuracy and BP scores for various levels of our feature augmentation method shows consistent trends among scores for a range of fairness measures, illustrating the benefit of the method for picking fairer models without significant loss of accuracy.

Increasing utilization of machine learning based decision support systems emphasizes the need for resulting predictions to be both accurate and fair to all stakeholders. In this work we present a novel approach to increase a Neural Network model's fairness during training. We introduce a family of fairness enhancing regularization components that we use in conjunction with the traditional binary-cross-entropy based accuracy loss. These loss functions are based on Bias Parity Score (BPS), a score that helps quantify bias in the models with a single number. In the current work we investigate the behavior and effect of these regularization components on bias. We deploy them in the context of a recidivism prediction task as well as on a census-based adult income dataset. The results demonstrate that with a good choice of fairness loss function we can reduce the trained model's bias without deteriorating accuracy even in unbalanced datasets.