Speech Referring Expression Comprehension (SREC) aims to localize the object in an image referred to by a spoken natural language query. However, raw speech is continuous and noisy, and prior ASR-free methods that align full utterances with transcripts using only global supervision can overfit to spurious correlations, limiting fine-grained grounding. To address this issue, we propose SliceCSRef, a robust SREC framework that improves generalization via dual-level semantic alignment. Beyond utterance-level speech–text alignment, SliceCSRef introduces slice-wise alignment that pairs randomly sampled speech segments with weakly matched transcript spans based on their relative temporal positions, providing fine-grained supervision without additional annotations. Experiments on six benchmarks show that SliceCSRef achieves state-of-the-art performance under standard settings and consistently improves robustness under truncated speech and playback-speed variations.

Motor imagery-based brain-computer interfaces (MI-BCIs) require abundant high-quality EEG data, yet collecting such data remains challenging. Recent diffusion probabilistic models (DPMs) have shown strong capabilities in generating realistic data, yet synthesizing EEG signals that preserve neuroscientific relevance remains challenging. To address this issue, we propose DiffEEGLossNet, a diffusion-based generative framework tailored for multi-channel EEG synthesis. By introducing an ERD/ERS-informed loss function, the model constrains the diffusion process to maintain motor imagery-specific neural patterns, enhancing both physiological plausibility and data fidelity. Experiments on the BCI Competition IV 2a and 2b datasets demonstrate that the generated EEG signals not only exhibit high realism and consistency but also significantly boost the performance of downstream EEG classification models through effective data augmentation.

With the widespread adoption of cloud computing and containerization technologies, large-scale clusters now host increasingly complex applications, yet the fragmentation of multidimensional resources (CPU, GPU, memory) has significantly degraded resource utilization efficiency. To address this challenge, this study investigates scheduling strategies for multidimensional resource fragmentation optimization. First, a unified quantitative metric is established by extending a task-statistics-based fragmentation measurement approach to threedimensional resource constraints (CPU, memory, GPU), enabling objective assessment of fragmentation levels. Subsequently, a multidimensional resource fragmentation optimization scheduling strategy is proposed based on the FGD framework. This approach employs TOPSIS to holistically address multidimensional resource fragmentation optimization. Building upon this methodology, we design and implement both two-dimensional and three-dimensional scheduling strategies. Comprehensive evaluation using an extended Kubernetes scheduler simulator with Alibaba production dataset reveals that: 1) All TOPSIS-based multi-dimensional strategies achieve performance comparable to or exceeding baseline approaches in their target dimensions while demonstrating synergistic effects.; 2) Two-dimensional strategies achieve optimal performance in their target dimensions but may degrade others 3) The three-dimensional optimization strategy TOPSIS-FGD-CGM can overcome the limitations of two-dimensional optimization strategies and demonstrates robust comprehensive performance.

Existing Kubernetes autoscaling solutions predominantly rely on single metrics with single-direction scaling (either horizontal or vertical), making it challenging to balance end-to-end latency and CPU utilization effectively. This paper introduces HyMetricScaler, a hybrid autoscaling framework. The framework employs a horizontal scaling component that makes decisions based on distributed tracing data and incorporates an exponential moving average mechanism to mitigate scaling oscillations, thereby optimizing end-to-end latency. Additionally, it utilizes a vertical scaling component based on PID control theory to adjust CPU quotas, improving CPU utilization. Experimental results demonstrate that HyMetricScaler significantly reduces oscillations in scaling decisions under high-load scenarios. In low-load scenarios, it explores the performance trade-off space by adjusting CPU target utilization. Through configuration adjustments of its components, the framework can achieve optimal end-to-end latency or CPU utilization, meeting various business requirements for different performance metrics. Comparative experiments against baseline algorithms validate the effectiveness of the proposed solution.

Since the advent of cloud computing, Elastic Computing (EC) has become the standard architecture for resource allocation and scheduling. EC typically allocates computing resources based on predefined specifications, such as virtual machine or container flavors. However, these flavors are often constrained by fixed CPU-to-memory ratios, which frequently fail to match the actual resource needs of applications. As a result, cloud providers experience high resource allocation rates nearing saturation ($> $80%) but with low utilization ($< $25%). This study introduces Flexible Computing (FC), a novel approach to resource allocation and scheduling. Unlike EC, FC allocates resources based on an application resource usage profile, derived from the historical resource consumption of workloads, rather than relying on fixed specifications. Additionally, FC incorporates a real-time performance degradation detection mechanism to address performance issues caused by the noisy-neighbor effect when colocated workloads interfere with each other. FC dynamically adjusts resource allocation according to actual usage, ensuring that application performance meets Service Level Agreements (SLAs), while preventing resource waste and performance degradation from improper resource over-commitment. Large-scale experimental validations conducted on the FC architecture within Huawei Cloud data centers demonstrate that, compared to EC, FC can reduce computing resource consumption by over 33% while managing the same workloads. Furthermore, FC's real-time performance degradation detection model achieves a prediction error of less than 5% across various testing environments, highlighting its commercial viability.

With the development of unmanned aerial vehicle (UAV) technology, UAV-enabled vehicular edge computing (VEC) has emerged as a powerful computational paradigm that improves edge resource efficiency. In particular, supporting stateful services, which require persistent context across offloading sessions, introduces new challenges. In the VEC systems, computation offloading, UAV trajectory planning, and service migration must be jointly optimized to maintain quality of service (QoS). However, existing works rarely consider this joint optimization, especially under high mobility scenarios. To fill this gap, this paper first considers the joint computation offloading, UAV trajectory, and service migration problem in the stateful service-oriented UAV-enabled VEC system and then formulates it as a dynamic multi-objective optimization problem, with the purpose of minimizing UAV flight cost, vehicle energy consumption, service migration time, and age of information (AoI). To effectively address the formulated problem, a novel joint Trajectory, Offloading, and Migration optimization approach (TOM) based on a dynamic multifactorial evolutionary algorithm is proposed. In particular, a service migration strategy is designed in TOM to efficiently migrate services in a parallel manner. In addition, an environmental adaptation strategy is triggered to cope with rapid dynamic changes in the environment. Extensive simulations on real-world datasets show that our proposed method outperforms several state-of-the-art peer methods.

Knowledge Tracing (KT) focuses on estimating students' knowledge states and predicting their future performances, which is a crucial task for online education platforms. In light of the advancements in educational big data and deep neural networks, numerous KT models have been proposed and promising outcomes have been achieved. Nevertheless, we have noted that current methods possess certain evident constraints. Thus, we propose a Knowledge Tracing model with Multi-Relational Integration (MRIKT): (1) we consider the more sophisticated relations between questions and skills, which can reveal deeper patterns of students' learning; (2) we emphasize the forgetfulness nature of students and the value of inter-exercises relations by incorporating a retentive module. Specifically, we choose graph convolutional networks to construct the advanced-relation between questions and skills, named graph representation module. Additionally, by linking different exercises, our novel retentive module, inspired by RetNet, can acquire valuable insights. We extensively evaluate the performance of MRIKT on three real-world datasets. The results demonstrate that MRIKT achieves outstanding performance, which improves at least 8.44% compared to baseline models.

Service providers tend to deploy application services to several different virtual machines (VMs) to improve the scalability and manageability of the cloud data center (CDC). Therefore, high frequency communication traffic is always involved among those VMs that are deployed the same application service. In order to reduce the communication cost (CC) of CDC, all VMs running the same service should be redeployed in the same subnet as much as possible by using live migration technology, because CC between VMs in different subnets is much higher than that within the same subnet. On the other hand, the migration time (MT) to complete all migration tasks is also crucial for providers and customers, because a prolonged MT will lead to the increased maintenance cost and the deterioration of quality of service (QoS). To address the aforementioned issues, this paper proposes an efficient service-aware VM scheduling approach (ES-VSA) based on a multi-objective evolutionary algorithm to minimize CC and MT, simultaneously. Finally, experiments are conducted on four different scenarios, and the simulation results demonstrate that our proposed algorithm is superior to several state-of-the-art algorithms in terms of both CC and MT.

Sharing learned models is crucial in research and the industry’s rapid development and progress. Meanwhile, as the Intellectual Property (IP) of the model proposer, the learned high-performance models must be protected to avoid being illegally copied or redistributed by malicious users. Unfortunately, even though the field of Electroencephalography (EEG) has made significant progress and the models are becoming increasingly complex, more work still needs to be done on protecting EEG-based models. The damage caused by model stealing and attack on the brain-computer interface (BCI) is more severe than in other fields. In this paper, we propose a method that protects the IP of EEG-based models with watermarking for the first time. Watermarks are embedded into three representative EEG-based models by designing a trigger set. On the premise of not sacrificing the primary task’s performance significantly, the models’ legality can be verified remotely through the trigger set. Furthermore, we demonstrate that the proposed model protection method is robust to various anti-watermarking attacks, such as fine-tuning, transfer learning, pruning, and watermark overwriting.

With the development of container technology, microservice architecture has become a powerful paradigm for cloud computing with efficient infrastructure management and large-scale service capabilities. Cloud providers require flexible resource management to meet dynamic workloads, such as autoscaling and provisioning. As one of the most popular open-source container orchestration systems, Kubernetes provides a built-in mechanism, Horizontal Pod Autoscaler (HPA), for dynamic resource autoscaling. However, the static rules of HPA are not adaptable to highly dynamic workloads. In this paper, we propose a deep reinforcement learning-based horizontal autoscaler(DScaler) for autoscaling of microservices deployed in Kubernetes. Under two workloads with different characteristics, our experiments show that the proposed approach reduces resource consumption by 19.90% and 10.80% while reducing SLA violations by 8.56% and 12.75% compared with HPA, respectively. In addition, our approach can significantly reduce resource consumption by about 60% compared to the existing reinforcement learning strategy while maintaining SLA within an acceptable level.

Collaborative Filtering (CF) is a prevalent technique in recommender systems.Substantial research focuses on learning the embedding of users and items via exploiting past user-item interactions.Recent years have witnessed the boom of Graph Convolutional Networks (GCNs) on CF.Performing graph convolution iteratively, GCN-based models concatenate/average/sum all outputs from different graph convolution layers to generate the embeddings of users and items.Although the previous methods have been proven effective, the pooling operations in the previous methods fail to consider the outputs from different graph convolution layers have different weights and the weights are related to sequential dependencies from precursor nodes.To resolve the aforementioned problems, in this work, we present a new model, Recurrent Neural Graph Collaborative Filtering (RNGCF), which proposes a sequential dependency construction module to adaptively generate the embeddings.Specifically, the module applies a gated recurrent unit (GRU) to learn the sequential dependencies from precursor nodes and an adaptive gated unit (AGU) to adaptively construct the embeddings based on the sequential dependencies.Extensive experiments on three benchmark datasets show that our model outperforms state-ofthe-art models consistently.Our implementation is available in PyTorch 1 .

Sequential recommendation aims to recommend the next item that a user will likely interact with by capturing the useful sequential patterns from users' historical behaviors.Recently, it has become an important and popular component in various e-commerce platforms.As a successful network, Transformer has been widely used to adaptively capture the dynamics of users' historical behaviors for sequential recommendation.In recommender systems, the size of embedding is usually set to be small.Under small embedding, the dot-product in Transformer may have the limitation on calculating the complex relevance between keys and queries.To address the common but neglected issue, in this paper, we present a new model, Deep Self-Attention for Sequential Recommendation (DSASrec), which proposes a chunking deep attention to compute attention weights.The chunking deep attention has two modules: a deep module and a chunking module.The deep module is used to improve the nonlinearity of the attention function.The chunking module is used to calculate attention weights several times like the multihead attention in Transformer.Extensive experiments on three benchmark datasets show that our model can achieve state-ofthe-art results.Our implementation is available in PyTorch 1 .

The large-scale video data on the web contain a lot of semantics, which are an important part of semantic web. Video descriptors can usually represent somewhat the semantics. Thus, they play a very important role in web multimedia content analysis, such as Scale-invariant feature transform (SIFT) feature. In this paper, we proposed a new video descriptor, called a temporal-compress and shorter SIFT(TC-S-SIFT) which can efficiently and effectively represent the semantics of web videos. By omitting the least discriminability orientation in three stages of standard SIFT on every representative frame, the dimensions of the shorter SIFT are reduced from 128-dimension to 96-dimension to save space storage. Then, the SIFT can be compressed by tracing SIFT features on video temporal domain, which highly compress the quantity of local features to reduce visual redundancy, and keep basically the robustness and discrimination. Experimental results show our method can yield comparable accuracy and compact storage size.

How to allocate workflow resources with cost constraint to optimize workflow time performance was studied. The average throughput time of workflow instances is used to measure the workflow time performance. A method based on queuing theory was proposed to calculate the average throughput time of workflow instances. Since the problem can be come down to an unbounded knapsack problem (UKP), an improved GA (genetic algorithm) suitable to solve the UKP was proposed to solve the problem. Examples were given to illustrate the feasibility and validity of the method. The results show that the algorithm has good evolution performance and can achieve or approach the optimal solutions. And compared with other allocation methods, our method performs best