Neural networks have achieved remarkable success in time series classification, but their reliance on large amounts of labeled data for training limits their applicability in cold-start scenarios. Moreover, they lack interpretability, reducing transparency in decision-making. In contrast, dynamic time warping (DTW) combined with a nearest neighbor classifier is widely used for its effectiveness in limited-data settings and its inherent interpretability. However, as a non-parametric method, it is not trainable and cannot leverage large amounts of labeled data, making it less effective than neural networks in rich-resource scenarios. In this work, we aim to develop a versatile model that adapts to cold-start conditions and becomes trainable with labeled data, while maintaining interpretability. We propose a dynamic length-shortening algorithm that transforms time series into prototypes while preserving key structural patterns, thereby enabling the reformulation of the DTW recurrence relation into an equivalent recurrent neural network. Based on this, we construct a trainable model that mimics DTW's alignment behavior. As a neural network, it becomes trainable when sufficient labeled data is available, while still retaining DTW's inherent interpretability. We apply the model to several benchmark time series classification tasks and observe that it significantly outperforms previous approaches in low-resource settings and remains competitive in rich-resource settings.

In traffic prediction, the goal is to estimate traffic speed or flow in specific regions or road segments using historical data collected by devices deployed in each area. Each region or road segment can be viewed as an individual client that measures local traffic flow, making Federated Learning (FL) a suitable approach for collaboratively training models without sharing raw data. In centralized FL, a central server collects and aggregates model updates from multiple clients to build a shared model while preserving each client's data privacy. Standard FL methods, such as Federated Averaging (FedAvg), assume that clients are independent, which can limit performance in traffic prediction tasks where spatial relationships between clients are important. Federated Graph Learning methods can capture these dependencies during server-side aggregation, but they often introduce significant computational overhead. In this paper, we propose a lightweight graph-aware FL approach that blends the simplicity of FedAvg with key ideas from graph learning. Rather than training full models, our method applies basic neighbourhood aggregation principles to guide parameter updates, weighting client models based on graph connectivity. This approach captures spatial relationships effectively while remaining computationally efficient. We evaluate our method on two benchmark traffic datasets, METR-LA and PEMS-BAY, and show that it achieves competitive performance compared to standard baselines and recent graph-based federated learning techniques.

Accurate estimation of Remaining Useful Life (RUL) and State of Health (SOH) is essential for Prognostics and Health Management (PHM) across a wide range of industrial applications. We propose a novel framework -- Reinforced Graph-Based Physics-Informed Neural Networks Enhanced with Dynamic Weights (RGPD) -- that combines physics-based supervision with advanced spatio-temporal learning. Graph Convolutional Recurrent Networks (GCRNs) embed graph-convolutional filters within recurrent units to capture how node representations evolve over time. Graph Attention Convolution (GATConv) leverages a self-attention mechanism to compute learnable, edge-wise attention coefficients, dynamically weighting neighbor contributions for adaptive spatial aggregation. A Soft Actor-Critic (SAC) module is positioned between the Temporal Attention Unit (TAU) and GCRN to further improve the spatio-temporal learning. This module improves attention and prediction accuracy by dynamically scaling hidden representations to minimize noise and highlight informative features. To identify the most relevant physical constraints in each area, Q-learning agents dynamically assign weights to physics-informed loss terms, improving generalization across real-time industrial systems and reducing the need for manual tuning. In both RUL and SOH estimation tasks, the proposed method consistently outperforms state-of-the-art models, demonstrating strong robustness and predictive accuracy across varied degradation patterns across three diverse industrial benchmark datasets.

Hacker forums provide critical early warning signals for emerging cybersecurity threats, but extracting actionable intelligence from their unstructured and noisy content remains a significant challenge. This paper presents an unsupervised framework that automatically detects, clusters, and prioritizes security events discussed across hacker forum posts. Our approach leverages Transformer-based embeddings fine-tuned with contrastive learning to group related discussions into distinct security event clusters, identifying incidents like zero-day disclosures or malware releases without relying on predefined keywords. The framework incorporates a daily ranking mechanism that prioritizes identified events using quantifiable metrics reflecting timeliness, source credibility, information completeness, and relevance. Experimental evaluation on real-world hacker forum data demonstrates that our method effectively reduces noise and surfaces high-priority threats, enabling security analysts to mount proactive responses. By transforming disparate hacker forum discussions into structured, actionable intelligence, our work addresses fundamental challenges in automated threat detection and analysis.

Low-dimensional chaotic systems such as the Lorenz-63 model are commonly used to benchmark system-agnostic methods for learning dynamics from data. Here we show that learning from noise-free observations in such systems can be achieved up to machine precision: using ordinary least squares regression on high-degree polynomial features with 512-bit arithmetic, our method exceeds the accuracy of standard 64-bit numerical ODE solvers of the true underlying dynamical systems. Depending on the configuration, we obtain valid prediction times of 32 to 105 Lyapunov times for the Lorenz-63 system, dramatically outperforming prior work that reaches 13 Lyapunov times at most. We further validate our results on Thomas' Cyclically Symmetric Attractor, a non-polynomial chaotic system that is considerably more complex than the Lorenz-63 model, and show that similar results extend also to higher dimensions using the spatiotemporally chaotic Lorenz-96 model. Our findings suggest that learning low-dimensional chaotic systems from noise-free data is a solved problem.

Driving trajectory data remains vulnerable to privacy breaches despite existing mitigation measures. Traditional methods for detecting driving trajectories typically rely on map-matching the path using Global Positioning System (GPS) data, which is susceptible to GPS data outage. This paper introduces CAN-Trace, a novel privacy attack mechanism that leverages Controller Area Network (CAN) messages to uncover driving trajectories, posing a significant risk to drivers' long-term privacy. A new trajectory reconstruction algorithm is proposed to transform the CAN messages, specifically vehicle speed and accelerator pedal position, into weighted graphs accommodating various driving statuses. CAN-Trace identifies driving trajectories using graph-matching algorithms applied to the created graphs in comparison to road networks. We also design a new metric to evaluate matched candidates, which allows for potential data gaps and matching inaccuracies. Empirical validation under various real-world conditions, encompassing different vehicles and driving regions, demonstrates the efficacy of CAN-Trace: it achieves an attack success rate of up to 90.59% in the urban region, and 99.41% in the suburban region.

This paper proposes a neural stochastic optimization method for efficiently solving the two-stage stochastic unit commitment (2S-SUC) problem under high-dimensional uncertainty scenarios. The proposed method approximates the second-stage recourse problem using a deep neural network trained to map commitment decisions and uncertainty features to recourse costs. The trained network is subsequently embedded into the first-stage UC problem as a mixed-integer linear program (MILP), allowing for explicit enforcement of operational constraints while preserving the key uncertainty characteristics. A scenario-embedding network is employed to enable dimensionality reduction and feature aggregation across arbitrary scenario sets, serving as a data-driven scenario reduction mechanism. Numerical experiments on IEEE 5-bus, 30-bus, and 118-bus systems demonstrate that the proposed neural two-stage stochastic optimization method achieves solutions with an optimality gap of less than 1%, while enabling orders-of-magnitude speedup compared to conventional MILP solvers and decomposition-based methods. Moreover, the model's size remains constant regardless of the number of scenarios, offering significant scalability for large-scale stochastic unit commitment problems.

Clinical monitoring of functional decline in ALS relies on periodic assessments that may miss critical changes occurring between visits. To address this gap, semi-supervised regression models were developed to estimate rates of decline in a case series cohort by targeting ALSFRS- R scale trajectories with continuous in-home sensor monitoring data. Our analysis compared three model paradigms (individual batch learning and cohort-level batch versus incremental fine-tuned transfer learning) across linear slope, cubic polynomial, and ensembled self-attention pseudo-label interpolations. Results revealed cohort homogeneity across functional domains responding to learning methods, with transfer learning improving prediction error for ALSFRS-R subscales in 28 of 32 contrasts (mean RMSE=0.20(0.04)), and individual batch learning for predicting the composite scale (mean RMSE=3.15(1.25)) in 2 of 3. Self-attention interpolation achieved the lowest prediction error for subscale-level models (mean RMSE=0.19(0.06)), capturing complex nonlinear progression patterns, outperforming linear and cubic interpolations in 20 of 32 contrasts, though linear interpolation proved more stable in all ALSFRS-R composite scale models (mean RMSE=0.23(0.10)). We identified distinct homogeneity-heterogeneity profiles across functional domains with respiratory and speech exhibiting patient-specific patterns benefiting from personalized incremental adaptation, while swallowing and dressing functions followed cohort-level trajectories suitable for transfer models. These findings suggest that matching learning and pseudo-labeling techniques to functional domain-specific homogeneity-heterogeneity profiles enhances predictive accuracy in ALS progression tracking. Integrating adaptive model selection within sensor monitoring platforms could enable timely interventions and scalable deployment in future multi-center studies.

However, due to their inherent sparsity, disorder, and lack of structure, instance-level semantic understanding of point clouds remains challenging - particularly under conditions of limited supervision and cross-modal semantic ambiguity. To address these issues, we propose SegV ec3D, a novel framework integrating attention mechanisms, embedding learning, and cross-modal alignment techniques for 3D point cloud instance segmentation. The proposed approach first builds a hierarchical instance feature extractor based on spatial adjacency and attention computation, enhancing the model's ability to capture fine-grained geometric structures. It then introduces a high-dimensional embedding space, enabling unsupervised instance segmentation through a contrastive-learning-based clustering mechanism. Furthermore, a shared cross-modal semantic space is constructed to align 3D data with natural language descriptions, allowing zero-shot understanding and retrieval of 3D objects given text queries. The model is ultimately deployed and validated in realistic scenarios, demonstrating strong generalizability and engineering feasibility. While recent methods like Mask3D [40] and ULIP [10][11] have advanced 3D segmentation and vision-language pre-training respectively, our approach uniquely integrates these domains by enabling instance segmentation with minimal labeling and directly aligning point clouds with language. Experimental evaluations confirm that the proposed method achieves high semantic discriminability, robust multi-modal alignment, and practical deployabil-ity. It supports weakly-supervised or unsupervised 3D instance understanding, providing a promising foundation for future multi-modal cognitive robotic systems.

In-context learning (ICL) has emerged as a powerful capability of large pretrained transformers, enabling them to solve new tasks implicit in example input-output pairs without any gradient updates. Despite its practical success, the mechanisms underlying ICL remain largely mysterious. In this work we study synthetic linear regression to probe how transformers implement learning at inference time. Previous works have demonstrated that transformers match the performance of learning rules such as Ordinary Least Squares (OLS) regression or gradient descent and have suggested ICL is facilitated in transformers through the learned implementation of one of these techniques. In this work, we demonstrate through a suite of out-of-distribution generalization experiments that transformers trained for ICL fail to generalize after shifts in the prompt distribution, a behaviour that is inconsistent with the notion of transformers implementing algorithms such as OLS. Finally, we highlight the role of the pretraining corpus in shaping ICL behaviour through a spectral analysis of the learned representations in the residual stream. Inputs from the same distribution as the training data produce representations with a unique spectral signature: inputs from this distribution tend to have the same top two singular vectors. This spectral signature is not shared by out-of-distribution inputs, and a metric characterizing the presence of this signature is highly correlated with low loss.