Traffic forecasting is fundamental to intelligent transportation systems, enabling congestion mitigation and emission reduction in increasingly complex urban environments. While recent graph neural network approaches have advanced spatial-temporal modeling, existing mixture-of-experts frameworks like Time-Enhanced Spatio-Temporal Attention Model (TESTAM) lack explicit incorporation of physical road network topology, limiting their spatial capabilities. We present TESTAM +, an enhanced spatio-temporal forecasting framework that introduces a novel SpatioSemantic Expert integrating physical road topology with data-driven feature similarity through hybrid graph construction. TESTAM + achieves significant improvements over TESTAM:$\mathbf{1. 3 \%}$MAE reduction on METR-LA ($\mathbf{3. 1 0}$vs. 3.14) and 4.1 % improvement on PEMS-BAY (1.65 vs. 1.72). Through comprehensive ablation studies, we discover that strategic expert selection fundamentally outperforms naive ensemble aggregation. Individual experts demonstrate remarkable effectiveness: the Adaptive Expert achieves$\mathbf{1. 6 3}$MAE on PEMS-BAY, outperforming the original three-expert TESTAM (1.72 MAE), while the SpatioSemantic Expert matches this performance with identical 1.63 MAE. The optimal Identity + Adaptive configuration achieves an$\mathbf{1 1. 5 \%}$MAE reduction compared to state-of-the-art MegaCRN on METR-LA (2.99 vs. 3.38), while reducing inference latency by 53.1 % compared to the full four-expert TESTAM+. Our findings reveal that fewer, strategically designed experts outperform complex multi-expert ensembles, establishing new state-of-the-art performance with superior computational efficiency for real-time deployment.

Bingzheng Gan, Yufan Zhao, Tianyi Zhang, Jing Huang, Li Yusu, Shu Xian Teo, Changwang Zhang, Wei Shi. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

ABSTRACT Vision Transformers (ViTs) have demonstrated strong capabilities in capturing global contextual information compared to convolutional neural networks, making them promising for remote sensing image analysis. However, ViTs often overlook critical local features, limiting their ability to accurately interpret intricate scenes. To address this issue, we propose an adaptive linear hybrid cross attention transformer (ALHCT). It integrates adaptive linear (AL) attention and hybrid cross (HC) attention to simultaneously learn local and global features. AL is introduced into ViT, as it helps reduce computational complexity from exponential to linear scale. Furthermore, ALHCT incorporates two adaptive linear swin transformers (ALST) to achieve multi‐scale feature representation, enabling the model to capture high‐level semantics and fine details. Finally, to enhance global perception and discriminative power, HC attention fuse local and global features which captured by the two ALST. Experiments on three remote sensing datasets demonstrate that ALHCT significantly improves classification accuracy, outperforming several state‐of‐the‐art methods, validating its effectiveness in classifying complex remote sensing scenes.

Large-scale clusters are often built with over-provisioned service resources, so as to satisfy the huge demand raised by enormous users in cloud environments. By estimating the resource demand of workloads, an on-demand resource provisioning method can be realized in these clusters, thus improving the energy efficiency. However, to guarantee Quality of Service (QoS), the resource demand of workload should be accurately estimated so as to provide suitable resources. Many statistical approaches estimate actual resource demand based on some workload features. But the relations between actual resource demand and workload features are generally obscure, and it's a big challenge to gain an accurate estimation under an obscure relation. In this paper, by considering a cluster as a queueing system, we construct a linearly dependent relation between resource demand and multiple feature combinations. The linearly dependent relation is inconstant due to its variable coefficients. Then, to ascertain specific relations which match actual situations, we design a Basic Linear regression (BL) algorithm. BL can obtain the optimal values for these coefficients, thus determining the inconstant relation to specific ones. Finally, we propose a Constructed Linear regression (CL) approach to estimate actual resource demands. CL forms a two-layer neural network by using several processes of BL as the neurons. To evaluate CL, we realize an On-Demand Resource Provisioning (ODRP) method in a typical power-aware cluster. Several evaluation metrics are proposed for conducting extensive experiments. The experimental results show that CL is effective to make accurate estimations.

We study multi-turn response generation for open-domain dialogues.The existing state-ofthe-art addresses the problem with deep neural architectures.While these models improved response quality, their complexity also hinders the application of the models in real systems.In this work, we pursue a model that has a simple structure yet can effectively leverage conversation contexts for response generation.To this end, we propose four auxiliary tasks including word order recovery, utterance order recovery, masked word recovery, and masked utterance recovery, and optimize the objectives of these tasks together with maximizing the likelihood of generation.By this means, the auxiliary tasks that relate to context understanding can guide the learning of the generation model to achieve a better local optimum.Empirical studies with three benchmarks indicate that our model can significantly outperform state-of-the-art generation models in terms of response quality on both automatic evaluation and human judgment, and at the same time enjoys a much faster decoding process.

While neural conversation models have shown great potentials towards generating informative and engaging responses via introducing external knowledge, learning such a model often requires knowledge-grounded dialogues that are difficult to obtain. To overcome the data challenge and reduce the cost of building a knowledge-grounded dialogue system, we explore the problem under a zero-resource setting by assuming no context-knowledge-response triples are needed for training. To this end, we propose representing the knowledge that bridges a context and a response and the way that the knowledge is expressed as latent variables, and devise a variational approach that can effectively estimate a generation model from independent dialogue corpora and knowledge corpora. Evaluation results on three benchmarks of knowledge-grounded dialogue generation indicate that our model can achieve comparable performance with state-of-the-art methods that rely on knowledge-grounded dialogues for training, and exhibits a good generalization ability over different datasets.

With the data growing explosively, data center networks (DCN) have to possess the characteristics of incrementally scalable, cost-efficient, high network capacity and fault tolerance. However, the widely used DCNs can not meet the demands above. In this paper, we propose a new type of data center topology named Criso to settle the challenges. Different from the existed works, Criso has the advantages of both switch-centric topologies (servers do not participate in routing) and the server-centric topologies (the scalability is not limited by the ports of switches). It is constructed based on pods, The internal structure of each pod is the same and there are only four external interfaces. By applying such structure, a pod-based and fault-tolerant routing algorithm is designed to handle multiple types of failures. Criso is hierarchically, recursively defined and high-network capacity which can scale up to millions of nodes. The analysis results demonstrate that the Criso model is significantly superior to four state-of-the-art data center structures in terms of the network capacity, scalability, cost, power consumption and other static characteristics. Criso achieves the target of low-cost, low-energy consumption and highly-scalability simultaneously.