Few-shot learning is critical in data mining systems where new or long-tail categories appear with only a few labeled examples. We propose MACL, a transductive and model-agnostic contrastive learning framework for robust few-shot classification. MACL contains two branches. The contrastive branch learns margin-aware embeddings through hard-negative mining, boundary-sensitive reweighting, and semi-positive construction. The transductive branch refines class weights by maximizing mutual information over the entire query set with support supervision. A lightweight gradient-based scheme adaptively balances the loss terms, and parameter-efficient tuning with adapters, LoRA, and visual prompts enables scalable adaptation across convolutional and transformer backbones. We construct controlled few-shot evaluations based on miniImageNet and tieredImageNet to assess robustness under spurious correlations, label noise, and open-set queries. Experiments show that MACL and its ViT-based variant consistently outperform strong inductive and transductive baselines in accuracy, calibration, and robustness. Overall, the results suggest that MACL benefits from a tightly coupled contrastive and transductive design in which margin-aware supervised contrastive learning shapes the posteriors optimized by the InfoMax objective under adaptive balancing. This cooperative mechanism goes beyond a simple combination of SupCon, TIM, and gradient reweighting and provides a unified and scalable solution for few-shot learning in practical data mining and knowledge discovery tasks.

Human activity recognition is a crucial task in healthcare and exercise sciences, with wide-ranging applications for monitoring human health and well-being, disease detection, and prevention. With a multitude of devices that record human activity proliferating people's daily lives in the form of smartwatches, various wearable devices, and cameras, the abundance of data collected has made it possible to apply neural architectures to the problem. This study considers human activity data collected as time series. Recent advancements in deep learning architectures, specifically the Transformer model, have made it possible to classify sequential data with excellent accuracy. This paper presents a human activity recognition system utilizing a combination of the OptiTrack PrimeX 22 motion capture camera and a Transformer-based classifier model.

Introduction: Positron Emission Tomography (PET) is a vital modality for investigating brain related disorders. However, data scarcity especially for novel molecular targets like neuroepigenetic enzymes combined with difficult-to-recruit patient populations limits the development of machine learning (ML) models. Our primary objective is to enhance single-subject classification of neuromolecular imaging data and facilitate biomarker discovery. We demonstrate our approach using histone deacetylase (HDAC) PET/MR imaging in Alcohol Use Disorder (AUD). Methods: We propose Catalysis Training pipeline, a framework that augments real imaging data with high-quality synthetic data generated by a Wasserstein Conditional Generative Adversarial Network (WCGAN). Using [11C]Martinostat PET/MR imaging, we extracted 1-D standardized uptake value ratio (SUVR) tabular features representing HDAC enzyme expression density across eight cingulate subregions. These were used to train and test ML classifiers, including Support Vector Machine (SVM), XGBoost, and Random Forest, under leave-one-out cross-validation. Results: Integrating synthetic data in the training process improved classification accuracy significantly: +26% for XGBoost and Random Forest (from 59% to 85%), and +18% for SVM (from 70% to 88%). Synthetic samples improved model generalizability. Key hemispheric and subregional cingulate HDAC patterns were also identified as potential biomarkers. Conclusion: Our results demonstrate that generative AI can help overcome data scarcity in low-data regime neuroimaging applications. Catalysis Training provides a scalable strategy to enhance ML-driven biomarker discovery and disease classification, especially for rare or difficult-to-study disorders like AUD. Clinically, cingulate HDAC expression measured by [11C]Martinostat PET/MR shows promise as an objective biomarker for AUD, complementing DSM-based diagnosis and informing novel treatment strategies.

Large Language Models (LLMs) have shown strong general capabilities in many applications. However, how to make them reliable tools for some specific tasks such as automated short answer grading (ASAG) remains a challenge. We present SteLLA (Structured Grading System Using LLMs with RAG) in which a) Retrieval Augmented Generation (RAG) approach is used to empower LLMs specifically on the ASAG task by extracting structured information from the highly relevant and reliable external knowledge based on the instructor-provided reference answer and rubric, b) an LLM performs a structured and question-answering-based evaluation of student answers to provide analytical grades and feedback. A real-world dataset that contains students' answers in an exam was collected from a college-level Biology course. Experiments show that our proposed system can achieve substantial agreement with the human grader while providing break-down grades and feedback on all the knowledge points examined in the problem. A qualitative and error analysis of the feedback generated by GPT4 shows that GPT4 is good at capturing facts while may be prone to inferring too much implication from the given text in the grading task which provides insights into the usage of LLMs in the ASAG system.

Deep learning models built upon Transformer architectures have led to substantial advancements in sequential data analysis. Nevertheless, their direct application to video-based tasks, such as Group Activity Recognition (GAR), remains constrained by the quadratic computational complexity and excessive memory requirements of global self-attention, especially when handling long video sequences. To overcome these limitations, we propose SUGAR: A Sequence Unfolding Based Transformer Model for Group Activity Recognition. Our approach introduces a novel sequence unfolding and folding mechanism that partitions long video sequences into overlapping local windows, enabling the model to concentrate attention within compact temporal regions. This local attention design dramatically reduces computational cost and memory footprint while maintaining high recognition accuracy. Within the Bi-Causal framework, SUGAR replaces conventional Transformer blocks, and experimental results on the Volleyball dataset demonstrate that our model achieves state-of-the-art performance, consistently exceeding 93% accuracy, with significantly improved efficiency. In addition, we investigate Lightning Attention 2 as an alternative linear-complexity attention module, identifying practical challenges such as increased memory usage and unstable convergence. To ensure robustness and training stability, we incorporate a dedicated safety mechanism that mitigates these issues. In summary, SUGAR offers a scalable, resource-efficient solution for group activity analysis in videos and exhibits strong potential for broader applications involving lengthy sequential data in computer vision and bioinformatics.

Accurately inferring latent orderings from unordered, high-dimensional tabular data is a core challenge in Big Data analytics, spanning domains from biology to finance, healthcare, and recommendation. We present FACT (Fusion-Augmented Causal Mask Transformer), a scalable framework that integrates domain-specific embeddings, gating-based fusion, and causality-inspired attention to infer interpretable trajectories under weak supervision. FACT combines pretrained language model embeddings with raw features, balances them via adaptive gating, and extends causal masking to permutation-invariant Set Transformers to model directional dependencies. To address supervision sparsity, we introduce a hybrid Sinkhorn-margin loss, aligning global orderings while enforcing local consistency. Experiments on biological and non-biological tabular datasets demonstrate that FACT improves ordering accuracy, interpretability, and computational efficiency compared to state-of-the-art baselines. While individual modules are based on established ideas, their integration delivers a practical, generalizable solution for ordering tasks across diverse Big Data applications.

Better model generalization is one key objective in machine learning. Although deep neural networks (DNNs) achieve impressive progress in many real-world applications, there is still a lack of fundamental understanding of why and when a DNN model generalizes well, which may hinder further improvement of DNN models. In this paper, we introduce a novel generalization measure, the Model Activation Ratio ( φ̅ ), which quantifies how broadly patterns and neurons of a DNN represent the training data. By measuring the average number of samples each neuron or pattern processes, φ̅ provides valuable insights into the interplay between memorization and generalization. In addition to quantifying generalization, we propose a training procedure called Progressive Neuron Rebirth (ProNeR), which integrates φ̅ directly into the training process. PRoNeR selectively reinitializes underutilized neurons to promote more balanced activation coverage, guiding the model toward improved generalization. We validate our approach with extensive experiments on several image benchmark datasets, using a variety of DNN architectures. The results show that incorporating φ̅ and PRoNeR leads to consistent and significant improvements in generalization performance.

The ACM KDD 2025 Health Day theme, "Harnessing AI Opportunities in Biomedicine and Healthcare" highlights the transformative potential of AI-driven applications in healthcare, translational biomedical research, and basic biological research. This extended abstract discusses recent advancements, challenges, and future directions, focusing on integrating AI-ready data sets, interdisciplinary collaborations, and ethical AI practices. It aims to catalyze discussions on the potential of AI ecosystems in revolutionizing healthcare and related fields.

Chaotic dynamics are ubiquitous in many real-world systems, ranging from biological and industrial processes to climate dynamics and the spread of viruses. These systems are characterized by high sensitivity to initial conditions, making it challenging to predict their future behavior confidently. In this study, we propose a novel deep-learning framework that addresses this challenge by directly exploiting the long-term compounding of local prediction errors during model training, aiming to extend the time horizon for reliable predictions of chaotic systems. Our approach observes the future trajectories of initial errors at a time horizon, modeling the evolution of the loss to that point through the use of two major components: 1) a recurrent architecture (Error Trajectory Tracing) designed to trace the trajectories of predictive errors through phase space, and 2) a training regime, Horizon Forcing, that pushes the model’s focus out to a predetermined time horizon. We validate our method on three classic chaotic systems and six real-world time series prediction tasks with chaotic characteristics. The results show that our approach outperforms the state-of-the-art methods.

Alcohol use disorder (AUD) is a disorder of clinical and public health significance requiring novel and improved therapeutic solutions. Both environmental and genetic factors play a significant role in its pathophysiology. However, the underlying epigenetic molecular mechanisms that link the gene-environment interaction in AUD remain largely unknown. In this proof-of-concept study, we showed, for the first time, the neuroepigenetic biomarker capability of non-invasive imaging of class I histone deacetylase (HDAC) epigenetic enzymes in the in vivo brain for classifying AUD patients from healthy controls using a machine learning approach in the context of precision diagnosis. Eleven AUD patients and 16 age- and sex-matched healthy controls completed a simultaneous positron emission tomography-magnetic resonance (PET/MR) scan with the HDAC-binding radiotracer [11C]Martinostat. Our results showed lower HDAC expression in the anterior cingulate region in AUD. Furthermore, by applying a genetic algorithm feature selection, we identified five particular brain regions whose combined [11C]Martinostat relative standard uptake value (SUVR) features could reliably classify AUD vs. controls. We validate their promising classification reliability using a support vector machine classifier. These findings inform the potential of in vivo HDAC imaging biomarkers coupled with machine learning tools in the objective diagnosis and molecular translation of AUD that could complement the current diagnostic and statistical manual of mental disorders (DSM)-based intervention to propel precision medicine forward.

This paper provides an overview of the ACM SIGKDD 2024 AI4Science4AI special day. It includes information about the organizers, invited speakers, keynote speakers, the event agenda, and insights from related workshops. The AI4Science4AI special day aims to bring together experts in artificial intelligence (AI) and science to discuss the latest developments, challenges, and future directions.

Mode collapse has been a persisting challenge in generative adversarial networks (GANs), and it directly affects the applications of GAN in many domains. Existing works that attempt to solve this problem have some serious limitations: models using optimal transport (OT) strategies (e.g., Wasserstein distance) lead to vanishing or exploding gradients; increasing the number of generators can cause several generators focusing on the same mode; and approaches that modify the loss also do not satisfactorily resolve mode collapse. In this article, we reduce mode collapse by formulating it as a Monge problem of OT map. We show that the Monge problem can be transformed to the distribution transformation problem in GAN, and a rectified affine neural network can be considered as a measurable function. In this way, we propose Monge GAN that uses this measurable function to transform the generated data distribution into the original data distribution. We utilize the Kantorovich formulation to obtain the OT cost, which is regarded as the OT distance between the two distributions. Finally, we conduct extensive experiments on both image and numerical datasets to validate our Monge GAN in reducing model collapse.

We address the challenging task of Universal Novel Category Discovery (UniNCD) in image classification, where models must distinguish between common and novel categories while avoiding the misclassification of novel categories as private-known ones. Previous prototype-based approaches face two major challenges: first, they significantly increase the negative transfer risk by often misaligning novel categories with private-known categories; second, they lead to sub-optimal prototypes because traditional prototype learning ignores diverse object characteristics of images, resulting in insufficient semantic guidance when optimizing instance representations using only instance-level prototypical distributions. To tackle these challenges, we present a Semantic-Enhanced Prototypical Network, dubbed SEPNet. This prototypical network is enhanced by refined prototypes and enriched semantics to learn better representations and avoid negative transfer, including three key ideas: (1) we design a Prototype Refinement (PR) strategy that can decouple common, private-known, and novel categories from unlabeled data, which can exclude misaligned prototypes to avoid negative transfer; (2) we attach prototypical distribution to each patch of images, which embed enhanced semantic information to prototypes and guide prototypical contrastive learning and, (3) we design a patch-entropy balance (PEB) method to encourage sparser patch-level prototypical distributions while maintaining the uniformity of dense distributions, sparsity emphasizes dominant category characteristics, and uniformity avoids the misguidance of irrelevant disturbance, thereby enhancing the distinctiveness of instances to the prototypes. Our method demonstrates superior performance on the UniNCD task across three benchmark datasets, outperforming existing state-of-the-art approaches by approximately 3.4% in terms of accuracy. We will release our code for reproduction.

Modern DNNs often include a huge number of parameters that are expensive for both computation and memory. Pruning can significantly reduce model complexity and lessen resource demands, and less complex models can also be easier to explain and interpret. In this paper, we propose a novel pruning algorithm, Cluster-Restricted Extreme Sparsity Pruning of Redundancy (CRESPR), to prune a neural network into modular units and achieve better pruning efficiency. With the Hessian matrix, we provide an analytic explanation of why modular structures in a sparse DNN can better maintain performance, especially at an extreme high pruning ratio. In CRESPR, each modular unit contains mostly internal connections, which clearly shows how subgroups of input features are processed through a DNN and eventually contribute to classification decisions. Such process-level revealing of internal working mechanisms undoubtedly leads to better interpretability of a black-box DNN model. Extensive experiments were conducted with multiple DNN architectures and datasets, and CRESPR achieves higher pruning performance than current state-of-the-art methods at high and extremely high pruning ratios. Additionally, we show how CRESPR improves model interpretability through a concrete example.

Time series data plays a significant role in many research fields since it can record and disclose the dynamic trends of a phenomenon with a sequence of ordered data points. Time series data is dynamic, of variable length, and often contains complex patterns, which makes its analysis challenging especially when the amount of data is limited. In this paper, we propose a multi-view feature construction approach that can generate multiple feature sets of different resolutions from a single dataset and produce a fixed-length representation of variable-length time series data. Furthermore, we propose a multi-encoder-decoder Transformer (MEDT) architecture to effectively analyze these multi-view representations. Through extensive experiments using multiple benchmarks and a real-world dataset, our method shows significant improvement over the state-of-the-art methods.

In this paper, we propose a novel task, Universal Category Discovery (UCD), to address the challenge of partial overlap between source and target domain categories. Different from previous tasks that assume all known categories exist in the target domain, UCD introduces "private-known" categories that only exist in the source domain and aims to classify unlabeled data as "common" or "novel" categories while avoiding misclassifying them into "private-known" categories. For this task, we propose a Tri-branch network with bidirectional Prototype-aware Matching (TriPM). TriPM effectively transfers knowledge from labeled to unlabeled data by bidirectionally matching similar data pairs, while a prototype matching strategy reduces the negative transfer risk from "private-known" categories. Finally, we propose a tri-branch network to decouple knowledge acquisition from labeled data, unlabeled data, and their interactions, which can avoid knowledge forgetting, explore novel patterns, and transfer common knowledge, respectively. Experiments demonstrate our model’s superiority over SOTA methods.

Analyzing animal behavior from video data is crucial for understanding brain function, assessing pharmacological interventions, and examining genetic modifications. Traditional methods often struggle to accurately analyze group behaviors in complex environments. To address these challenges, we introduce the Animal Joint Embedded Prediction Architecture (Animal-JEPA), a novel self-supervised learning model designed for studying animal behavior from video data. Animal-JEPA leverages a dynamic scaling mechanism and an elliptical masking strategy to enhance feature extraction and behavioral analysis without the need for labeled data. Our approach significantly outperforms existing models, including Separate 3D ConvNet (S3D) [3], Video Vision Transformer (ViViT) [4], and the original V-JEPA [6], particularly in multi-category and multi-objective classification tasks on our newly developed Mice-Behavior3 (MB3) dataset. The results highlight Animal-JEPA’s potential to improve the accuracy and adaptability of behavioral analysis in animal research, providing a powerful tool for neuroscientists and researchers.

The seismic fortification intensity requirements of buildings of various heights are different. Identifying the heights of different buildings using a specific method can reduce secondary disasters such as building damage and house collapse caused by destructive earthquakes. The traditional building object detection model uses the feature expression of the dataset based on manually sketching the shape of the building in the remote sensing images, which will cause less acquired image information, unclear target features of buildings, and the decrease of target detection accuracy of identifying building types. The real-time target detection algorithm YOLO has fast detection speed and high accuracy, but its main disadvantage is the inaccurate positioning of the building bounding box and the misjudgment of partially overlapping building objects. Image feature retrieval using large data set domain adaptive fast algorithm is a new research hotspot in current image recognition. This research proposes a multi-branch network with the addition of a module using attention mechanism to coordinate the salient and sub-salient information of images. The algorithm adds channel and spatial attention mechanisms to the feature extraction network, and uses weighting and filtering methods to perform residual fusion on the original features vectors in images. The model is deployed on the remote sensing geographic information evaluation platform to conduct various building target detection experiments. The average detection accuracy of the urban buildings can reach 81.8%, and the detection speed for remote sensing segmentation images is 26 per second. The mAP@0.5 of the YOLOv5 algorithm with the attention mechanism is 12% higher than that of the unimproved YOLOv5 algorithm on the urban street view image dataset.

Effective early warning systems for extreme flood events in large river basins necessitate reliable long-lead streamflow forecasts. However, the inherent uncertainty within each phase of the weather system-rainfall prediction, runoff generation, and streamflow prediction-amplifies with each stage, rendering accurate long-lead streamflow estimations challenging. In response to this, our study introduces a novel deep-learning-based model, the Cascaded Spatio-Temporal Learning Deep Network (CASTLE). CASTLE synergistically integrates observed upstream precipitation, recent streamflow data, and short-term precipitation forecasts derived from a selection of quantitative climate models to produce an accurate streamflow estimate. Specifically, we employ deep residual architectures on both observed and forecasted precipitation data to model the cascading spatio-temporal processes, which begin with upstream rainfall, move to rainfall-runoff, and finally conclude with downstream discharge. Our aim is to identify hidden space-time patterns that can be used to forecast future downstream flow over extended periods. We assess CASTLE’s efficacy by forecasting the downstream discharge of the Ganges River over a long lead time. Results show that our approach outperforms the current state-of-the-art streamflow forecasting models.