Real-world multivariate time series often exhibit multi-scale temporal dynamics and intricate inter-variable dependencies, making long-term forecasting particularly challenging. In this work, we propose a global token-driven multi-scale forecasting framework with dual-attention fusion. To capture multi-scale periodic patterns, the input sequence is segmented into multi-scale patches based on candidate periods, with the dominant ones derived via fast Fourier transform. Global tokens are then introduced as shared representations to integrate information from the temporal and variable dimensions, and a multi-scale patch-token interaction module is designed to establish interactions between the patches and global tokens, enabling the capture and aggregation of temporal dependencies across different scales. A dual-attention fusion module, employing both self-attention and cross-attention mechanisms, is then proposed to capture intrinsic and context-aware variable correlation among variables. To integrate cross-variable and cross-scale information into patch representations, a global information fusion module is designed. Finally, a period-aware weighting approach is devised to adaptively fuse the multi-scale predictions. Comprehensive experiments demonstrate that the proposed framework achieves state-of-the-art performance across various real-world datasets.

Accurate and timely estimation of tropical cyclone (TC) wind radii is significant for characterizing wind structure, disaster prevention, and mitigation. The existing methods have not sufficiently considered and utilized multimodal (i.e., multisource heterogeneous) data for wind radii estimation. Meanwhile, complete modalities (i.e., all used modalities) can hardly be available simultaneously, especially in real-time monitoring scenarios, which restricts the applicability of multimodal estimation models. It is challenging to maintain the accuracy of wind radii estimates when confronted with the issue of missing-modality. Therefore, to address these issues, this article aims to achieve robust TC wind radii estimation under both conditions with complete modalities and missing-modality. We first present a multimodal fusion network, MT-TCNet, for estimating TC wind radii under conditions with complete modalities. MT-TCNet benefits from multimodal data including satellite infrared (IR) images, reanalysis of wind fields, and the physical parameter maximum sustained wind (MSW) speed. MSW, which reflects TC intensity, is incorporated to embed the implicit relationship between TC intensity and wind radii. It is capable of providing superior and robust wind radii estimates in scenarios without time constraints, and can be used to generate long-term historical results. Furthermore, this article proposes MT-TCNet-Distill to alleviate the issue of missing-modality caused by delays in ERA5 reanalysis wind fields through generalized distillation and missing modality imputation. MT-TCNet-Distill broadens the applicability of MT-TCNet, which heavily relies on reanalysis data, enabling robust wind radii estimation in real-time scenarios. Comprehensive experiments demonstrate the superior performance of MT-TCNet and MT-TCNet-Distill compared to state-of-the-art methods.

With the explosive development of e-commerce for service, tens of millions of orders are generated every day on the Meituan platform. By allocating bonuses to new customers when they pay, the Meituan platform encourages them to use its own payment service for a better experience in the future. It can be formulated as a multi-choice knapsack problem (MCKP), and the mainstream solution is usually a two-stage method. The first stage is user intent detection, predicting the effect for each bonus treatment. Then, it serves as the objective of the MCKP, and the problem is solved in the second stage to obtain the optimal allocation strategy. However, this solution usually faces the following challenges: (1) In the user intent detection stage, due to the sparsity of interaction and noise, the traditional multi-treatment effect estimation methods lack interpretability, which may violate the domain knowledge that the marginal gain is non-negative with the increase of the bonus amount in economic theory. (2) There is an optimality gap between the two stages, which limits the upper bound of the optimal value obtained in the second stage. (3) Due to changes in the distribution of orders online, the actual cost consumption often violates the given budget limit. To solve the above challenges, we propose a framework that consists of three modules, i.e., User Intent Detection Module, Online Allocation Module, and Feedback Control Module. In the User Intent Detection Module, we implicitly model the treatment increment based on deep representation learning and constrain it to be non-negative to achieve monotonicity constraints. Then, in order to reduce the optimality gap, we further propose a convex constrained model to increase the upper bound of the optimal value. For the third challenge, to cope with the fluctuation of online bonus consumption, we leverage a feedback control strategy in the framework to make the actual cost more accurately approach the given budget limit. Finally, we conduct extensive offline and online experiments, demonstrating the superiority of our proposed framework, which reduced customer acquisition costs by 5.07% and is still running online.

The interference in Device-to-Device (D2D) communications system is a major challenge. To cope with the severe interference, interference management techniques such as power control are needed. Because of the ability to learn automatically from the environment, the Q-learning algorithm has already been used as the D2D power control technique in many previous studies. But the algorithm is very time-consuming because of its multiple iterations. In this paper, a D2D power control method based on Hierarchical Extreme Learning Machine (H-ELM) is proposed. H-ELM is an effective supervised learning algorithm evolved from original Extreme Learning Machine (ELM). By comparing with the other two power control algorithms based on machine learning, distributed Q-learning and CART Decision Tree, the simulation results show that the method in this paper has a better performance in both communication throughput and energy efficiency with limited time consumption.

In this paper, reinforcement learning (RL) based power control algorithm in underlay D2D communication is studied. The approach we use regards D2D communication as a multi-agents system, and power control is achieved by maximizing system capacity while maintaining the requirement of quality of service(QoS) from cellular users. We propose two RL based power control methods for D2D users, i.e., team-Q learning and distributed-Q learning. The former is a centralized method in which only one Q-value table needs to be maintained, while the latter enables D2D users to learn independently and reduces the complexity of Q-value table. Simulation results show the difference of the two Q-learning algorithm in terms of convergence and reward function. In addition, it is shown that through our distributed-Q learning, D2D users not only are able to learn their power in a self-organized way, but also achieve better system performance than that using traditional method in LTE(Long Term Evolution).

There are three types of web service QoS criterion weight: subjective weight decided by user's preference of QoS criterion, objective weight affected by QoS criterion value and synthetic weight synthesized with the two. Considering the uncertainty of conversion from qualitative value, which is used to describe user's preference, to quantitative value, the uncertainty of subjective weight is described by cloud model and the weight is calculated using normalization method and least square method. Objective weight is obtained using entropy method. For synthetic weight, Cobb-Douglas method is used to synthesize subjective and objective weight. Lastly, a case is studied and sensibility analysis of weight based on uncertain weight is discussed.