Abstract
Traffic flow prediction remains a complex task due to the intricate spatial and temporal correlations in real-world traffic data. Although existing graph neural network (GNN) approaches have shown promise in capturing these relationships, their high computational requirements limit their suitability for real-time deployment. To overcome these limitations, we propose spatiotemporal adaptive refinement with knowledge distillation (STARK), a novel and efficient framework that integrates graph fusion with adaptive knowledge distillation (AKD) in a spatiotemporal graph convolutional network (STGCN). Our method leverages graph fusion to capture both localized and global traffic dynamics, enhancing adaptability across diverse traffic conditions. It further employs two dedicated teacher models that independently emphasize spatial and temporal features, guiding a lightweight student model through a distillation process that dynamically adjusts based on prediction uncertainty. This adaptive learning mechanism enables the student model to prioritize and better learn from more difficult prediction instances. Evaluations on four benchmark traffic datasets [PEMS03, PEMS04, PEMSD7(M), and PEMS08] demonstrate that STARK achieves competitive predictive performance, measured by mean absolute error (MAE) and root mean square error (RMSE), while significantly reducing computational overhead. Our approach thus offers an effective and scalable solution for real-time traffic forecasting.

Abstract
Large Events Models (LEMs) are a class of models designed to predict and analyze the sequence of events in soccer matches, capturing the complex dynamics of the game. The original LEM framework, based on a chain of classifiers, faced challenges such as synchronization, scalability issues, and limited context utilization. This paper proposes a unified and scalable approach to model soccer events using a tabular autoregressive model. Our models demonstrate significant improvements over the original LEM, achieving higher accuracy in event prediction and better simulation quality, while also offering greater flexibility and scalability. The unified LEM framework enables a wide range of applications in soccer analytics that we display in this paper, including real-time match outcome prediction, player performance analysis, and game simulation, serving as a general solution for many problems in the field.

Abstract
[No abstract available]

Abstract
Traditional recurrent neural networks operate as passive observers of data, unable to modify the information they learn from despite errors that may arise from suboptimal input representations. We introduce Read & Write LSTM (read-write LSTM), a new variant within the family of read & write machine learning (RW-ML) architectures that address this fundamental limitation by integrating input modification directly into the backpropagation process. Read-write LSTM establishes a dynamic feedback loop where input representations evolve alongside model weights through gradient transformation mechanisms. Our approach introduces a principled gradient scaling framework with an adaptive correction rate that carefully controls the extent of data modification, preserving data integrity while enhancing representational power. We comprehensively evaluate read-write LSTM against traditional LSTMs and state-of-the-art transformer models on the M4 competition and Numenta Anomaly Benchmark datasets, demonstrating significant improvements in forecasting accuracy. Notably, read-write LSTM consistently out-performs standard LSTM models in over 70% of time series with complex patterns and achieves superior performance on 55% of anomaly-rich datasets. Through extensive experimentation and analysis, we establish both the theoretical foundations and practical benefits of integrating data modification with neural computation, paving the way for a new generation of adaptive learning systems that actively reshape their inputs rather than merely adapting to them.

Abstract
Identifying the types of transportation modes that people use is a central problem in transportation research. Effective feature construction plays a crucial role in developing a successful machine learning model. In this study, we demonstrate an approach to identify commuters' transportation modes solely using raw GPS trajectory data. First, we transform the representation of location data points into a vector of motion features in the time domain. Next, we create fixed-length instances in the time domain. Subsequently, we transform the instances time-domain features into frequency-domain features using the fast Fourier transform. This results in a pool of features for the instances in both the time and frequency domains. We use the Sequential Forward Floating Selection technique to select the most informative features to train our models. We evaluate our approach using two distinct real-world GPS trajectory datasets. Our results show that the random forest classifier achieved an ROC-AUC scores of 79% and 89% on the respective datasets.

Abstract