High-risk event prone driver identification considering driving behavior temporal covariate shift

Drivers who perform frequent high-risk events (e.g., hard braking maneuvers) pose a significant threat to traffic safety. Existing studies commonly estimated high-risk event occurrence probabilities based upon the assumption that data collected from different time periods are independent and identically distributed (referred to as i.i.d. assumption). Such approach ignored the issue of driving behavior temporal covariate shift, where the distributions of driving behavior factors vary over time. To fill the gap, this study targets at obtaining time-invariant driving behavior features and establishing their relationships with high-risk event occurrence probability. Specifically, a generalized modeling framework consisting of distribution characterization (DC) and distribution matching (DM) modules was proposed. The DC module split the whole dataset into several segments with the largest distribution gaps, while the DM module identified time-invariant driving behavior features through learning common knowledge among different segments. Then, gated recurrent unit (GRU) was employed to conduct time-invariant driving behavior feature mining for high-risk event occurrence probability estimation. Moreover, modified loss functions were introduced for imbalanced data learning caused by the rarity of high-risk events. The empirical analyses were conducted utilizing online ride-hailing services data. Experiment results showed that the proposed generalized modeling framework provided a 7.2% higher average precision compared to the traditional i.i.d. assumption based approach. The modified loss functions further improved the model performance by 3.8%. Finally, benefits for the driver management program improvement have been explored by a case study, demonstrating a 33.34% enhancement in the identification precision of high-risk event prone drivers.

Exploring Regularization Methods for Domain Generalization in Accelerometer-Based Human Activity Recognition

The study of Domain Generalization (DG) has gained considerable momentum in the Machine Learning (ML) field. Human Activity Recognition (HAR) inherently encompasses diverse domains (e.g., users, devices, or datasets), rendering it an ideal testbed for exploring Domain Generalization. Building upon recent work, this paper investigates the application of regularization methods to bridge the generalization gap between traditional models based on handcrafted features and deep neural networks. We apply various regularizers, including sparse training, Mixup, Distributionally Robust Optimization (DRO), and Sharpness-Aware Minimization (SAM), to deep learning models and assess their performance in Out-of-Distribution (OOD) settings across multiple domains using homogenized public datasets. Our results show that Mixup and SAM are the best-performing regularizers. However, they are unable to match the performance of models based on handcrafted features. This suggests that while regularization techniques can improve OOD robustness to some extent, handcrafted features remain superior for domain generalization in HAR tasks.

Hierarchical resampling for bagging in multistudy prediction with applications to human neurochemical sensing

We propose the "study strap ensemble", which combines advantages of two common approaches to fitting prediction models when multiple training datasets ("studies") are available: pooling studies and fitting one model versus averaging predictions from multiple models each fit to individual studies. The study strap ensemble fits models to bootstrapped datasets, or "pseudo-studies." These are generated by resampling from multiple studies with a hierarchical resampling scheme that generalizes the randomized cluster bootstrap. The study strap is controlled by a tuning parameter that determines the proportion of observations to draw from each study. When the parameter is set to its lowest value, each pseudo-study is resampled from only a single study. When it is high, the study strap ignores the multi-study structure and generates pseudo-studies by merging the datasets and drawing observations like a standard bootstrap. We empirically show the optimal tuning value often lies in between, and prove that special cases of the study strap draw the merged dataset and the set of original studies as pseudo-studies. We extend the study strap approach with an ensemble weighting scheme that utilizes information in the distribution of the covariates of the test dataset. Our work is motivated by neuroscience experiments using real-time neurochemical sensing during awake behavior in humans. Current techniques to perform this kind of research require measurements from an electrode placed in the brain during awake neurosurgery and rely on prediction models to estimate neurotransmitter concentrations from the electrical measurements recorded by the electrode. These models are trained by combining multiple datasets that are collected in vitro under heterogeneous conditions in order to promote accuracy of the models when applied to data collected in the brain. A prevailing challenge is deciding how to combine studies or ensemble models trained on different studies to enhance model generalizability. Our methods produce marked improvements in simulations and in this application. All methods are available in the studyStrap CRAN package.

ROBUST WHITE MATTER HYPERINTENSITY SEGMENTATION ON UNSEEN DOMAIN

Typical machine learning frameworks heavily rely on an underlying assumption that training and test data follow the same distribution. In medical imaging which increasingly begun acquiring datasets from multiple sites or scanners, this identical distribution assumption often fails to hold due to systematic variability induced by site or scanner dependent factors. Therefore, we cannot simply expect a model trained on a given dataset to consistently work well, or generalize, on a dataset from another distribution. In this work, we address this problem, investigating the application of machine learning models to unseen medical imaging data. Specifically, we consider the challenging case of Domain Generalization (DG) where we train a model without any knowledge about the testing distribution. That is, we train on samples from a set of distributions (sources) and test on samples from a new, unseen distribution (target). We focus on the task of white matter hyperintensity (WMH) prediction using the multi-site WMH Segmentation Challenge dataset and our local in-house dataset. We identify how two mechanically distinct DG approaches, namely domain adversarial learning and mix-up, have theoretical synergy. Then, we show drastic improvements of WMH prediction on an unseen target domain.