Time-Series (TS) exhibits pronounced non-stationarity. Consequently, most forecasting methods display compromised robustness to concept drift, despite the prevalent application of instance normalization. We tackle this challenge by first analysing concept drift through a bias-variance lens and proving that weighted ensemble reduces variance without increasing bias. These insights motivate DeepBooTS, a novel end-to-end dual-stream residual-decreasing boosting method that progressively reconstructs the intrinsic signal. In our design, each block of a deep model becomes an ensemble of learners with an auxiliary output branch forming a highway to the final prediction. The block‑wise outputs correct the residuals of previous blocks, leading to a learning‑driven decomposition of both inputs and targets. This method enhances versatility and interpretability while substantially improving robustness to concept drift. Extensive experiments, including those on large-scale datasets, show that the proposed method outperforms existing methods by a large margin, yielding an average performance improvement of 15.8% across various datasets, establishing a new benchmark for TS forecasting.

This paper focuses on the energy-efficient unmanned aerial vehicles (UAVs) assisted vehicular networks, where multiple rotary-wing UAVs are deployed to provide uplink service to numerous ground vehicles. The objective is to maximize the long-term system energy efficiency (EE) over the entire service period, through jointly optimizing the vehicle-UAV association, the sub-carrier assignment, the power control of vehicles and the trajectory design of UAVs. The formulated problem involves the limitations on the transmit power of vehicles and the propulsion power of UAVs, the quality of service (QoS) requirements of vehicles, the UAV movement constraints, and co-channel interference among vehicles. This makes the problem a mixed integer non-convex fractional programming problem accompanied by a mass of variables and diverse constraints, which is difficult to be solved within the polynomial time through traditional optimization methods. To cope with the timely decision-making requirement and dynamic moving scenario, we opt to the deep reinforcement learning (DRL) approach. To do so, the formulated problem is first transformed to a Markov decision process (MDP). Then, an improved k-means proximal policy optimization (IKPP) algorithm is proposed to solve the MDP problem. The proposed algorithm involves action reconstruction, the improved k-means algorithm, and proximal policy optimization-clip algorithm, which can help obtain the solutions with low complexity. Simulation results demonstrate the convergence, scalability and real-time of the proposed algorithm, along with its performance advantage over other benchmark algorithms.

The number of categories of instances in the real world is normally huge, and each instance may contain multiple labels. To distinguish these massive labels utilizing machine learning, eXtreme Label Classification (XLC) has been established. However, as the number of categories increases, the number of parameters and nonlinear operations in the classifier also rises. This results in a Classifier Computational Overload Problem (CCOP). To address this, we propose a Multi-Head Encoding (MHE) mechanism, which replaces the vanilla classifier with a multi-head classifier. During the training process, MHE decomposes extreme labels into the product of multiple short local labels, with each head trained on these local labels. During testing, the predicted labels can be directly calculated from the local predictions of each head. This reduces the computational load geometrically. Then, according to the characteristics of different XLC tasks, e.g., single-label, multi-label, and model pretraining tasks, three MHE-based implementations, i.e., Multi-Head Product, Multi-Head Cascade, and Multi-Head Sampling, are proposed to more effectively cope with CCOP. Moreover, we theoretically demonstrate that MHE can achieve performance approximately equivalent to that of the vanilla classifier by generalizing the low-rank approximation problem from Frobenius-norm to Cross-Entropy. Experimental results show that the proposed methods achieve state-of-the-art performance while significantly streamlining the training and inference processes of XLC tasks. The source code has been made public at https://github.com/Anoise/MHE.

In this article, we focus on the problem of product defect detection (PDD) in air conditioner (AC) manufacturing. The challenges are twofold: first, the scale of the objects undergoes significant variations, thereby increasing the difficulty of detection; second, the computing power of terminals is limited, and it may be difficult to meet the quasi-real-time detection requirements. Therefore, to improve detection accuracy and speed, a lightweight object detection model tailored for deployment on the compact wireless Raspberry Pi, is proposed: a deep cascaded U-shape network is presented to effectively capture both global context and local details of objects, which can reduce the feature redundancy and the number of the model parameters. An adaptive multiscale squeeze-and-excitation is designed for feature reuse and fusion, enhancing both detection accuracy and efficiency. Then, to meet the quasi-real-time detection demands and better utilize end-edge-cloud resources in industrial internet of things (IIoTs), an actor–critic-based dynamic offloading (ACDO) algorithm is proposed to minimize the long-term cumulative time of task detection. ACDO utilizes the elapsed time as the reward to directly optimize the mixed variables of the task to achieve efficient offloading. The proposed methods are verified at an AC manufacturing line, which demonstrates accurate and quasi-real-time defect detection, achieving a 64% reduction in runtime and a 1.2% improvement in average mean average precision. In addition, we publish two PDD datasets to accelerate the related research.

This work investigates proactive edge caching for device-to-device (D2D)-assisted wireless networks, where user equipment (UE) can be selected as caching nodes to assist content delivery to reduce the content transmission latency. In doing so, there are two challenges: 1) how to precisely get the user’s preference to cache the proper contents at UEs and 2) how to replace the contents cached at UEs when there are new popular contents emerging. To address these, we develop a user preference learning-based proactive edge caching (UPL-PEC) strategy. In the strategy, we first propose a novel context and social-aware user preference learning method to precisely predict user’s dynamic preferences by jointly exploiting the context correlation among different contents, the influence of social relationships and the time-sequential patterns of user’s content requests. Specifically, the bidirectional long short-term memory networks are adopted to capture the time-sequential patterns of the user’s content requests. And, the graph convolutional networks are developed to capture the high-order similarity representation among different contents from the constructed content graph. To learn the social influence representation, an attention mechanism is designed to generate the social influence weights to users with different social relationship. Based on the learned user preference, a proactive edge caching architecture is proposed to integrate the offline caching content placement and the online caching content replacement policy to continuously cache the popular contents at UEs. Simulation results show that the proposed UPL-PEC strategy outperforms the existing similar caching strategies at about 3.13%–4.62% in terms of the average content transmission latency.

Wireless traffic prediction has drawn increasing research interests as it can provide guidance to the network optimization. With the predicted information, one can preassign the resources on demand and perform network congestion control adaptively. The network efficiency is therefore enhanced. However, the wireless traffic prediction in the context of mobile scenario, such as Internet of Vehicles (IoVs), is still a challenge issue. The mobile nature of devices, which dynamically changes the topology of network, would brings difficulties to the prediction. This paper focuses on the deep learning based wireless traffic prediction in the IoVs scenario. We first propose a novel method to match up the movement- and communication-behavior of users, by merging two independent datasets on the trajectories of vehicles and communication traffic volumes together. Then a novel STeP-UNet is proposed, in which the SpatioTemporal Partial (STeP) Convolutional Neural Network module is embedded to capture cross-domain features of the wireless traffic pattern, and the UNet structure is utilized to realize the skipping connection from front layer to back layer to fuse different resolutions. Experimental results confirms the promising performance of the proposed model, where 4%~8% performance improvement over other benchmark methods can be achieved.

This study explores whether neural networks can classify multiple samples simultaneously in a forward process. Therefore, a multi‐input multi‐prediction network architecture has been proposed. The authors call this method a multi‐sample inference network (MSIN). In addition to maximising the use of network shared parameters, the network can also use multiple samples for training. MSIN allows multiple samples to be randomly combined to act as data augmentation, and the random combination of corresponding labels can regularise the network as a loss regularisation, which makes MSIN have better generalisation performance. In contrast, category expansion is a problem that is difficult to solve because neural networks can only predict a fixed number of categories. The network proposed in this study can solve the category expansion problem by expanding the initial layers and the final layers. It is trained by using samples of multiple domains at the same time to ensure that the network has no significant decline in the predictive performance of the existing categories. The MSIN method can also be applied to the generative adversarial network to enable it to simultaneously generate samples of multiple sample domains.