Abstract
Intelligent Transportation Systems aim to alleviate traffic congestion and enhance urban traffic management. Transformer-based methods have shown promise in traffic prediction due to their capability to handle long-range dependencies. However, they disregard local context during parallel processing and can be computationally expensive for large traffic networks. On the other hand, they miss the hierarchical information hidden in regions of large traffic networks. To address these issues, we introduce CSCN, a novel framework that clusters traffic sensors based on data similarity, employs clustered multi-head self-attention for efficient hierarchical pattern learning, and utilizes causal convolutional attention for capturing local temporal trends. In addition to these advancements, we integrate snapshot ensemble learning into CSCN, allowing for the exploitation of diverse snapshots obtained during training to enrich predictive performance. Evaluations of real-world data highlight CSCN's superiority in traffic flow prediction, showcasing its potential for enhancing transportation systems with improved accuracy and efficiency.

Abstract
Taxi demand prediction with scarce historic information is among the most encountered challenges of the present decade for the traffic network of a smart city. Lack of sufficient information results in the failure of conventional approaches in prediction for a new city. Additionally, the prevalent Deep Neural Network (DNN) Models resort to ineffectual approaches which fail to meet the required prediction performance for the network. Moreover, existing domain adaptation (DA) models could not sufficiently reap the domain-shared features well from multiple source, questioning the models' applicability. Complex structure of these DA models tends to a nominal performance gain due to inefficient resource utilization of the sources. The present paper introduces a domain adaptation deep neural network model, Bootstrap Zero-Shot (BTS-Z) learning model which focuses on capturing the latent spatio-temporal features of the whole city traffic network shared among every source city and maneuver them to predict for the target city traffic network with no prior information. The presented model proves the efficacy of the bootstrap algorithm in the prediction of demands for the unseen target over the computationally expensive MAML models. The experimental results on three real-world city taxi data on the standard benchmark metrics report a minimum of 23.41% improvement over the best performing competitive system. © 2024 IEEE.

Abstract
Accurate identification of transportation mode distribution is essential for effective urban planning. Recent advancements in machine learning have spurred research on automated Transportation Mode Detection (TMD). While existing TMD methods predominantly employ standard flat classification methods, this paper introduces HiClass4MD, a novel hierarchical approach. By leveraging the misclassification errors from standard flat classifier, HiClass4MD learns the class hierarchy for transportation modes. Although hierarchical metrics initially indicated performance improvements when applied to real-world GPS trajectories dataset, a subsequent evaluation using conventional metrics revealed inconsistent results. While decision trees benefited marginally, other classifiers exhibited no significant gains or even degraded. This study highlights the complexity of applying hierarchical classification to TMD and underscores the need for further investigation into the factors influencing its effectiveness.

Abstract
In many organizations with a distributed operation, not only is data collection distributed, but models are also developed and deployed separately. Understanding the combined knowledge of all the local models may be important and challenging, especially in the case of a large number of models. The automated development of consensus models, which aggregate multiple models into a single one, involves several challenges, including fidelity (ensuring that aggregation does not penalize the predictive performance severely) and completeness (ensuring that the consensus model covers the same space as the local models). In this paper, we address the latter, proposing two measures for geometrical and distributional completeness. The first quantifies the proportion of the decision space that is covered by a model, while the second takes into account the concentration of the data that is covered by the model. The use of these measures is illustrated in a real-world example of academic management, as well as four publicly available datasets. The results indicate that distributional completeness in the deployed models is consistently higher than geometrical completeness. Although consensus models tend to be geometrically incomplete, distributional completeness reveals that they cover the regions of the decision space with a higher concentration of data.

Abstract
Self-adaptive Systems (SaS) are becoming increasingly important for adapting to dynamic environments and for optimizing performance on resource-constrained devices. A practical approach to achieving self-adaptability involves using a Pareto-Front (PF) to store the system's hyper-parameters and the outcomes of hyperparameter combinations. This paper proposes a novel method to approximate a PF, offering a configurable number of solutions that can be adapted to the device's limitations. We conducted extensive experiments across various scenarios, where all PF solutions were replaced, and real world scenarios were performed using actual measurements from a Human Activity Recognition (HAR) system. Our results show that our method consistently outperforms previous methods, mainly when the maximum number of PF solutions is in the order of hundreds. The effectiveness of our method is most apparent in real-case scenarios where it achieves, when executed in a Raspberry Pi 5, up to 87% energy consumption reduction and lower execution times than the second-best algorithm. Additionally, our method ensures a more evenly distributed solution across the PF, preventing the high concentration of solutions. © 2024 IEEE.

Abstract
This study investigates the impact of class imbalance and its potential interplay with other factors on machine learning models for transportation mode classification, utilising two real-world GPS trajectory datasets. A Random Forest model serves as the baseline, demonstrating strong performance on the relatively balanced dataset but experiencing significant degradation on the imbalanced one. To mitigate this effect, we explore various state-of-the-art class imbalance learning techniques, finding only marginal improvements. Resampling the fairly balanced dataset to replicate the imbalanced distribution suggests that factors beyond class imbalance are at play. We hypothesise and provide preliminary evidence for class overlap as a potential contributing factor, underscoring the need for further investigation into the broader range of classification difficulty factors. Our findings highlight the importance of balanced class distributions and a deeper understanding of factors such as class overlap in developing robust and generalisable models for transportation mode detection.