The proliferation of smartphone technology has generated unprecedented volumes of data, creating challenges in under-standing digital behavior patterns. We present a new computational system that integrates Large Language Models (LLMs) with conventional data processing techniques using a novel three-levelOur system employs LLMs for zero-shot learning capabilities to classify usage patterns, achieving 95% accuracy in task classification through automated pattern identification. The system implements pattern verification reaching 98% validation accuracy and utilizes automated validation that reduces data loss by 75%. This hierarchical approach demonstrates consistent performance across diverse device types and usage scenarios while maintaining processing efficiency through automated prompt engineering and code generation.

Predicting and quantifying the impact of traffic accidents is necessary and critical to Intelligent Transport Systems (ITS). As a state-of-the-art technique in graph learning, current graph neural networks heavily rely on graph Fourier transform, assuming homophily among the neighborhood. However, the homophily assumption makes it challenging to characterize abrupt signals such as traffic accidents. Our paper proposes an abrupt graph wavelet network (AGWN) to model traffic accidents and predict their time durations using only one single snapshot.

Globally, traffic accidents account for over 3,700 daily deaths, equating to 1.35 million deaths annually. Studies show that collision avoidance systems can significantly reduce the probability and intensity of accidents. Time-to-accident (TTA) is considered the principal parameter for collision avoidance systems allowing for decision-making in traffic, dynamic path planning, and accident mitigation. Despite the importance of TTA, the literature has insufficient research on TTA estimation for traffic scenarios. The majority of recent work focuses on accident anticipation by providing a probabilistic measure of an immediate or future collision. We propose to forecast TTA based on Spatio-temporal features extracted from accident videos obtained via dashboard cameras. Our model can also recognize accident and non-accident scenes with 100% accuracy. Additionally, the impact of spatial resolution and temporal depth on prediction error is analyzed in this work. We implement state-of-the-art video learning architectures and compare the results against static image architectures. Our comprehensive experiments suggest that leveraging Spatio-temporal features is an effective method to estimate TTA. Our best model can estimate the TTA with an average prediction error of 0.30 seconds with a mean prediction horizon of 3.4 seconds.

Approximately 2,000 satellites orbiting Earth relay telecommunications, broadcasting, and data communications to and from different locations globally. For example, in 2018,8.4 million households re-lied on satellite internet in the United States. However, extreme space weather, space environment phenomena driven by plasma be-tween stars and planets, can harm satellites and the global data and internet communications, and affect near-Earth space. Extracting features that represents those plasma waves requires highly-trained space scientists. We want to design machine learning (ML) meth-ods to automatically design features and extract information from plasma waves for the early detection of extreme space weather. To do that and to leverage the rich and state-of-the-art algorithms in computer vision, we first use Heliophysics Audified: Resonances in Plasmas (HARP) Sonification Data Processing package [1] to con-vert magnetospheric Ultra-Low Frequency (ULF) waves to sound and spectrograms. We then utilize unsupervised learning meth-ods to cluster plasma waves into different groups to capture the commonalities and differences between those activities. This initial and pilot exploration in the field offers the potential of practical applications of ML to space science field. The results will help with satellite-based internet and communications as part of the edge computing community.

Sound is an important modality to perceive and understand the spatial environment. With the development of digital technology, massive amounts of smart devices in use around the world can collect sound data. Auditory spatial scenes, a spatial environment to understand and distinguish sound, are important to be detected by analyzing sounds collected via those devices. Given limited annotated auditory spatial samples, the current best-performing model can predict an auditory scene with an accuracy of 73%. We propose a novel yet simple Sliding Window based Convolutional Neural Network, SlideCNN, without manually designing features. SlideCNN leverages windowing operation to increase samples for limited annotation problems and improves the prediction accuracy by over 12% compared to the current best-performing models. It can detect real-life indoor and outdoor scenes with a 85% accuracy. The results will enhance practical applications of ML to analyze auditory scenes with limited annotated samples. It will further improve the recognition of environments that may potentially influence the safety of people, especially people with hearing aids and cochlear implant processors.

Electroencephalograph (EEG) data is used to diagnose brain conditions, such as epilepsy. The brain gives off electrical activity in voltages at different parts of the cerebral cortex. When electroen-cephalograph (EEG) data is taken, analyzing the data can show which part of the brain has activity and how much activity. However, currently studies only consider spatial and temporal parts of brain activities separately. In this study, we propose to fuse spatio-temporal information together via imitation learning to better understand brain activities, especially cortical speech-evoked responses. We will validate our methods via a real-life dataset to understand the patterns and distinguish hearing-impaired individuals from normal-hearing individuals based on brain activities (i.e., cortical speech-evoked responses). To the best of our knowledge, we are the first group to use imitation learning for brain activity study, especially the cortical speech-evoked responses. Our methods have the potential to be integrated as a sustainable service and can be leveraged for future hearing research.

Predicting and measuring the impact of traffic collisions is crucial for Intelligent Transportation Systems (ITS). Numerous works in this field have successfully applied graph neural networks to ITS. Existing research on graph neural networks mainly relies on the graph Fourier transform, assuming neighborhood homophily. The homophily assumption, on the other hand, makes it difficult to define abrupt signals such as traffic accidents. Our research proposes an abrupt graph wavelet network (AGWN) for forecasting the durations of traffic incidents using a single shot. To begin, graph wavelet (GW) is theoretically examined in terms of linear separability in comparison to graph Fourier (GF), demonstrating its advantage in modeling abrupt graph signals. Sensitivity analysis and admissibility conditions are utilized to further study the behavior of GW in abrupt graph signals, justifying the use of zero sum function as wavelet kernel. The synthetic data results support our proposed wavelet kernel's effectiveness in modeling a variety of abrupt signals, while real-world trials demonstrate that our method significantly outperforms baseline models in forecasting the duration of an accident impact.

Effective training of advanced ML models requires large amounts of labeled data, which is often scarce in scientific problems given the substantial human labor and material cost to collect labeled data. This poses a challenge on determining when and where we should deploy measuring instruments (e.g., in-situ sensors) to collect labeled data efficiently. This problem differs from traditional pool-based active learning settings in that the labeling decisions have to be made immediately after we observe the input data that come in a time series. In this paper, we develop a real-time active learning method that uses the spatial and temporal contextual information to select representative query samples in a reinforcement learning framework. To reduce the need for large training data, we further propose to transfer the policy learned from simulation data which is generated by existing physics-based models. We demonstrate the effectiveness of the proposed method by predicting streamflow and water temperature in the Delaware River Basin given a limited budget for collecting labeled data. We further study the spatial and temporal distribution of selected samples to verify the ability of this method in selecting informative samples over space and time.

In order to meet the health needs of the coming "age wave," technology needs to be designed that supports remote health monitoring and assessment. In this study we design clinician in the loop (CIL), a clinician-in-the-loop visual interface, that provides clinicians with patient behavior patterns, derived from smart home data. A total of 60 experienced nurses participated in an iterative design of an interactive graphical interface for remote behavior monitoring. Results of the study indicate that usability of the system improves over multiple iterations of participatory design. In addition, the resulting interface is useful for identifying behavior patterns that are indicative of chronic health conditions and unexpected health events. This technology offers the potential to support self-management and chronic conditions, even for individuals living in remote locations.