Depression represents a prevalent and costly mental health issue, presenting significant challenges for screening of a large population. Traditional diagnostic methods depend on clinical interviews that are time-consuming, resource-intensive, and require clinician time, limiting accessibility. While recent machine learning approaches utilizing interview modalities show promise, they frequently struggle with the small dataset sizes common in mental health research. To address this, we investigate the effectiveness of deploying multi-task learning frameworks whose primary task of classification is boosted by jointly training with an auxiliary task in a multi-task manner. Specifically, we leverage an auxiliary task of masked value imputation alongside the primary depression classification task applied to the DAIC-WOZ dataset. Using extracted facial feature information in conjunction with existing recurrent and attention-based time series imputation methods, our results demonstrate that the multi-task framework achieves strong classification performance, with balanced accuracy scores reaching up to 0.94 on individual core dataset questions. These findings suggest both the viability of automated screening tools using privacy-preserving features and the potential for shortened interview times by focusing on the most discriminative questions, thereby reducing patient burden while maintaining screening effectiveness.

The pressures faced by college students frequently lead to heightened levels of stress. Wearable devices, which collect sensor data in a non-intrusive manner, present an opportunity for early detection of stress. Nonetheless, there is a lack of diversity in current research concerning psychological assessments, physiological metrics, and time series features. In this work, we utilize a Fitbit dataset and evaluate its use in predicting stress through machine learning. Our results demonstrate that physiological data such as calories burned and sleep hold promise for stress screening, with F1 scores reaching up to 0.81. These findings illustrate the potential of wearable technology for continuous stress monitoring and emphasize the need for selecting appropriate data aggregation levels and physiological modalities for effective screening.

Major depressive disorder (MDD) and post-traumatic stress disorder (PTSD) are mental disorders that reduce quality of life. As they are challenging to detect in a timely manner, recent studies explore the mental illness screening potential of language models on modalities such as transcripts. Such datasets suffer from a limited number of participants. To overcome these challenges, we take a two-pronged approach: (1) we leverage multi-task learning to model multiple illnesses concurrently namely MDD and PTSD screening, and (2) we plug in pre-trained language models as a backbone, namely, Bidirectional Encoder Representations from Transformers (BERT) variants – which are capable of learning the linguistic content of clinical interview transcripts. In particular, we experiment with three multi-task weighting strategies and five BERT variants, applying them to 15 transcript sets extracted from the Distress Analysis Interview Corpus. Our results indicate that leveraging multi-task learning, especially with meta-weighting, increases the screening performance compared to single-task learning. Our multi-task learning model strategies improved the F1 scores for all 15 datasets for both MDD and PTSD screening. Notably, multitask learning improved MDD screening ability by 20 percent with transcripts regarding ‘regret’ and achieved F1 of 0.89 and 0.82 respectively for MDD and PTSD screening with transcripts regarding ‘medical history’. Our findings may help researchers develop more effective mental illness screening models.

Generative modeling is a powerful technique that involves creating machine learning models capable of creating new data similar to the data it was trained on. Generative Adversarial Networks (GANs) are a leading approach for generative modeling. However, GAN training is known to be a notoriously difficult task. GAN convergence issues are largely caused by the supports of the real and generated distributions being disjoint. To tackle this open problem, we propose a novel GAN pre-training process that effectively aligns the supports of the generated and real data prior to applying traditional adversarial GAN training. The key component of our method, called AlignGAN, is learning a mapping between the input data distribution and a latent representation defined over a hypersphere, regularized by a One Class Classifier. This successfully encourages the generator to produce samples throughout the support of the real data, while not generating samples outside the support. We maintain support alignment through low-bandwidth noise convolutions and additional One Class regularization, leading to continued stable GAN training. We validate our approach against leading stabilization methods on three benchmark datasets, showing AlignGAN routinely produces the best results.

Generative models are being increasingly used in science and industry applications. Unfortunately, they often perpetuate the biases present in their training sets, such as societal biases causing certain groups to be underrepresented in the data. For instance, image generators may overwhelmingly produce images of white people due to few non-white samples in their training data. It is imperative to debias generative models so they synthesize an equal number of instances for each group, while not requiring retraining of the model to avoid prohibitive expense. We thus propose a *distribution mapping module* that produces samples from a *fair noise distribution*, such that the pretrained generative model produces *semantically uniform* outputs - an equal number of instances for each group - when conditioned on these samples. This does *not* involve retraining the generator, nor does it require *any* real training data. Experiments on debiasing generators trained on popular real-world datasets show that our method outperforms existing approaches.

Abstract Text classification is a common task in data science. Despite the superior performances of deep learning based models in various text classification tasks, their black‐box nature poses significant challenges for wide adoption. The knowledge‐to‐action framework emphasizes several principles concerning the application and use of knowledge, such as ease‐of‐use, customization, and feedback. With the guidance of the above principles and the properties of interpretable machine learning, we identify the design requirements for and propose an interpretable deep learning (IDeL) based framework for text classification models. IDeL comprises three main components: feature penetration, instance aggregation, and feature perturbation. We evaluate our implementation of the framework with two distinct case studies: fake news detection and social question categorization. The experiment results provide evidence for the efficacy of IDeL components in enhancing the interpretability of text classification models. Moreover, the findings are generalizable across binary and multi‐label, multi‐class classification problems. The proposed IDeL framework introduce a unique iField perspective for building trusted models in data science by improving the transparency and access to advanced black‐box models.