Sequential recommendation leverages interaction sequences to predict forthcoming user behaviors, crucial for crafting personalized recommendations. However, the true preferences of a user are inherently complex and high-dimensional, while the observed data are merely a simplified and low-dimensional projection of the rich preferences, which often leads to prevalent issues like data sparsity and inaccurate model training. To learn true preferences from the sparse data, most existing works endeavor to introduce some extra information or design some ingenious models. Although they have shown to be effective, extra information usually increases the cost of data collection, and complex models may result in difficulty in deployment. Innovatively, we avoid the use of extra information or alterations to the model; instead, we fill the transformation space between the observed data and the underlying preferences with randomness. Specifically, we propose a novel model-agnostic and highly generic framework for sequential recommendation called sample enrichment via temporary operations on subsequences (SETO), which temporarily and separately enriches the transformation space via sequence enhancement operations with rationality constraints in training. The transformation space not only exists in the process from input samples to preferences but also in preferences to target samples. We highlight our SETO’s effectiveness and versatility over multiple representative and state-of-the-art sequential recommendation models (including six single-domain sequential models and two cross-domain sequential models) across multiple real-world datasets (including three single-domain datasets, three cross-domain datasets and a large-scale industry dataset).

In cloud computing, static resource configurations create a trade-off: tenants overprovision to avoid resource starvation, causing inefficiency and cost, while providers suffer low utilization despite high allocations. To improve efficiency, providers often use overcommitted environments where multiple workloads share hosts, but this leads to interference and potential Quality-of-Service (QoS) violations. This paper introduces a real-time dynamic control framework that mitigates interference by adaptively pinning workloads to CPU groups. Using deep reinforcement learning (DRL) with the Proximal Policy Optimization (PPO) algorithm, an intelligent agent continuously adjusts CPU pinning based on realtime feedback to maintain Service-Level-Agreement (SLA) compliance. Experiments under two optimization objectives-overallperformance-first and priority-performance-first-show that the proposed approach improves overall QoS by approximate to 25% compared with static pinning. When prioritization is enabled, high-priority workloads gain significant performance improvements while lower-priority ones remain within SLA limits. These results demonstrate that a DRL-based CPU-pinning strategy effectively manages resource contention in overcommitted clouds, enhancing utilization while upholding tenant SLAs.

Accurate cluster-level resource forecasting is pivotal for cloud capacity planning. Existing static heuristics fail to adapt to dynamic, heterogeneous workloads, causing resource waste or service disruptions. This paper proposes ASP (Adaptive Scaling Predictor), a learning-based framework optimized for cloud scenarios via three core innovations: a dual-path input fusing long-term periodic and short-term real-time features to capture global and local temporal patterns comprehensively; a MoE-Transformer architecture that enhances learning capacity and cross-scenario transferability, with sparse activation reducing deployment resource demands; a business-driven loss function balancing resource savings and task queuing rates, prioritizing service reliability alongside utilization. Experiments on Huawei Cloud and Alibaba Cloud production traces show ASP outperforms expert heuristics and state-of-the-art models, achieving 40.23% resource savings in Huawei Cloud dataset, a 0.01% task queuing rate, and a 10% underestimation ratio (vs. 20% for iTransformer) in Alibaba Cloud dataset.

Heterogeneous sequential recommendation (HSR) aims to learn dynamic behavior dependencies from the diverse behaviors of user-item interactions to facilitate precise sequential recommendation. Despite many efforts yielding promising achievements, there are still challenges in modeling heterogeneous behavior data. One significant issue is the inherent sparsity of a real-world data, which can weaken the recommendation performance. Although auxiliary behaviors (e.g., clicks) partially address this problem, they inevitably introduce some noise, and the sparsity of the target behavior (e.g., purchases) remains unresolved. Additionally, contrastive learning-based augmentation in existing methods often focuses on a single behavior type, overlooking fine-grained user preferences and losing valuable information. To address these challenges, we have meticulously designed a behavior-aware dual-channel preference learning framework (BDPL). This framework begins with the construction of customized behavior-aware subgraphs to capture personalized behavior transition relationships, followed by a novel cascade-structured graph neural network to aggregate node context information. We then model and enhance user representations through a preference-level contrastive learning paradigm, considering both long-term and short-term preferences. Finally, we fuse the overall preference information using an adaptive gating mechanism to predict the next item the user will interact with under the target behavior. Extensive experiments on three real-world datasets demonstrate the superiority of our BDPL over the state-of-the-art models.

Band selection (BS) is a key method in hyperspectral image (HSI) classification that helps to reduce the computational burden and improve the class separability. However, with the emerging of uncrewed aerial vehicle (UAV)-borne HSI datasets, their attributes, such as high spatial and spectral resolution as well as large-scale samples pose serious challenges to the existing BS methods, making them inefficient. In addition, the efficient utilization of the prior knowledge from the data collected by fixed UAV-borne sensors in different regions is often easily overlooked. In view of these issues, this article proposes a neural network-assisted particle swarm optimization (PSO) algorithm for cross-domain BS of UAV-borne HSIs. First, a knowledge learning strategy is designed for the source domain, which applies a neural network model to learn the useful prior knowledge in labeled source domain data. Then, a network-assisted PSO algorithm is introduced to search for the optimal subset of bands in the target domain under the guidance of the valid prior knowledge captured from the source domain by the network model. Moreover, a similarity-based grouping strategy is designed to group similar bands and then select bands from each group with the aims of reducing the redundant information in the subset of bands. Finally, experimental results on three common UAV-borne HSI datasets show that our proposed method can efficiently handle UAV-borne HSI data with large samples, as it is able to find a subset of bands with higher quality compared to several state-of-the-art BS methods.

Gradual typing effectively combines the advantages of dynamic and static typing and is utilized in Information Flow Control. The gradual guarantee ensures that removing type annotations does not affect runtime behavior. However, Toro et al. identify a tension between the gradual guarantee and noninterference, leading many gradual security-typed languages to reconcile these two properties. Recently, Chen and Siek’s language $\lambda_{\text {IFC }}^{\star}$ achieved both properties. Nevertheless, their cast calculus suffers from scalability issues and lacks binary operators, which limits its practical use.In this paper, we present a partial cast calculus $\lambda^{p}$ for gradual information security based on the concept of partial label. Partial labels serve as concise representations for security coercions and runtime labels, simplifying the definition of label-related operators and facilitating cast reductions within the semantics of $\lambda^{p}$. Consequently, our calculus effectively supports binary operators and naturally supports richer security lattices. We formally prove both noninterference and the gradual guarantee for $\lambda^{p}$. Furthermore, we present a gradual surface language $\lambda^{g}$ along with a compilation from surface terms to cast-calculus terms. We prove that this compilation preserves typing, precision relations, and observational equivalence. Finally, we extend our calculus to accommodate unknown references and provide a semantic interpretation of partial labels.

While previous works on Multi-Behavior Sequential Recommendation (MBSR) have made great efforts to exploit the behavior types, they often fail to fully leverage the behavior and attribute information to learn different aspects of user preferences. Firstly, for predicting user preferences under target behaviors, there is considerable room for improvement by capturing both intra- and inter-behavior item transition relationships at the sequence level, as well as extracting the inter-behavior common and different aspects of a user’s preferences at each timestep based on the contextual sequential information. Secondly, in MBSR, the effect of behavior types on item attributes has not been revealed, which is very important in real-world applications since a behavior such as view representing a user’s weak preferences could not explicitly indicate the user’s interests on an attribute. To address the above two issues, we propose a novel solution called Behavior-aware Attribute-infused Sequential Encoders (BASE) . Specifically, a module named tri-path preference learning is designed to capture user multi-aspect interests via a sequence encoder, a behavior encoder and an attribute encoder. Besides, a behavior-aware dual-granularity contrastive learning module constructs sample pairs in item and attribute granularities to enhance the model’s ability to utilize potential item and attribute information. Extensive empirical studies on three public datasets show that our BASE significantly outperforms various state-of-the-art methods from five different categories. The datasets and our implementation codes are released at https://github.com/Erin-Gr/BASE .

Large language models have transformed AI-assisted software engineering, but current research remains biased toward high-resource languages such as Python, with weaker performance in languages like Rust and OCaml. Since real-world systems are inherently polyglot, robust multilingual code intelligence is crucial. This survey focuses on two key tasks: multilingual code generation from shared natural-language requirements, and multilingual code translation that preserves semantics across languages. It reviews representative methods, benchmarks, and evaluation metrics, and highlights challenges and opportunities for trustworthy cross-language generalization.

Multi-agent large language model systems can tackle complex multi-step tasks by decomposing work and coordinating specialized behaviors. However, current coordination mechanisms typically rely on statically assigned roles and centralized controllers. As agent pools and task distributions evolve, these design choices can lead to inefficient routing, poor adaptability, and fragile fault recovery. We introduce Symphony-Coord, a task-local coordination framework with decentralized execution that transforms agent selection into an online multi-armed bandit problem. Instead of relying on a fixed task-to-role map, Symphony-Coord allows routing specializations to emerge from interaction and feedback. The framework employs a two-stage dynamic beacon protocol:(i) a lightweight candidate screening mechanism to limit communication and computation overhead; and (ii) an adaptive LinUCB selector that routes subtasks using context features derived from task requirements and agent states, updated through delayed post-execution feedback. Under candidate-conditional linear bandit assumptions, we prove sublinear regret bounds for the immediate-feedback selector and explicitly separate the deferred-update effects introduced by post-vote rewards. Validation through simulation experiments and real-world large language model benchmarks shows that Symphony-Coord improves task routing efficiency and recovery behavior under distribution shifts and agent failures.

Early cancer detection using serum biomarkers is clinically valuable but limited by insufficient sensitivity in single-marker approaches and the lack of interpretability in existing machine learning models. We propose MTL-DNN, an attention-enhanced multi-task deep learning framework that simultaneously performs binary cancer detection and multi-class cancer type classification from serum biomarkers. This architecture incorporates a shared encoder for cross-task feature extraction, a dual-branch self-attention mechanism to dynamically weigh biomarkers in each task, and a SHAP-based interpretability module per patient. On two practical datasets (1,818 and 627 samples) MTL-DNN achieves AUC $=0.94$ and accuracy $=0.91$ for binary detection, Macro-F1 $=0.87$ with Top-2 accuracy $=0.96$ across seven-class cancer types, significantly outperforming XGBoost and single-task DNN baselines ($\mathrm{p}<$ 0.001). SHAP analysis identifies clinically validated biomarkers including CA-125, CEA and AFP. These results demonstrate MTL-DNN as an accurate and interpretable framework for non-invasive cancer screening.

In media cloud environments, shared instances often suffer from unpredictable QoS degradation, leading to impaired user experience in real-time multimedia services such as HD video streaming. Existing methods struggle to model the high-dimensional, nonlinear telemetry dynamics that precede performance drops. To address this, we propose MTS-CR, a novel framework that combines contrastive learning and representation learning for real-time QoS degradation detection. MTS-CR employs a Linear Attention-based Encoder enhanced by Contrastive Predictive Coding (CPC) to extract discriminative temporal representations from 73 system-level metrics (CPU, memory, I/O, network). A lightweight MLP-based Predictor then forecasts imminent QoS anomalies. Evaluated on real-world cloud telemetry, MTS-CR significantly outperforms state-of-the-art baselines in both detection accuracy and generalization, offering cloud operators an actionable signal for dynamic resource orchestration — ultimately preserving end-user QoE.

Multi-behavior recommendation (MBR) aims to improve the performance w.r.t. the target behavior (i.e., purchase) by leveraging auxiliary behaviors (e.g., click, favourite). However, in real-world scenarios, a recommendation method often needs to process different types of behaviors and generate personalized lists for each task (i.e., each behavior type). Such a new recommendation problem is referred to as multi-behavior multi-task recommendation (MMR). So far, the most powerful MBR methods usually model multi-behavior interactions using a cascading graph paradigm. Although significant progress has been made in optimizing the performance of the target behavior, it often neglects the performance of auxiliary behaviors. To compensate for the deficiencies of the cascading paradigm, we propose a novel solution for MMR, i.e., behavior-informed graph embedding learning (BiGEL). Specifically, we first obtain a set of behavior-aware embeddings by using a cascading graph paradigm. Subsequently, we introduce three key modules to improve the performance of the model. The cascading gated feedback (CGF) module enables a feedback-driven optimization process by integrating feedback from the target behavior to refine the auxiliary behaviors preferences. The global context enhancement (GCE) module integrates the global context to maintain the user's overall preferences, preventing the loss of key preferences due to individual behavior graph modeling. Finally, the contrastive preference alignment (CPA) module addresses the potential changes in user preferences during the cascading process by aligning the preferences of the target behaviors with the global preferences through contrastive learning. Extensive experiments on two real-world datasets demonstrate the effectiveness of our BiGEL compared with ten very competitive methods.

Social influence plays a pivotal role in understanding collective decision-making, as individuals readily observe and are influenced by their friends’ behaviours. Accurately predicting social influence at the individual level is essential for various applications, including political communication and marketing. Graph Neural Networks (GNNs) have been widely applied for modelling social influence. However, existing work often overlooks higher-order interactions and hierarchical structures. It also struggles to effectively handle the multi-type nature of social networks. In this study, we propose a framework that learns latent feature representations of users in an end-to-end manner to forecast social influence. A critical component of the framework is a novel geometric deep learning paradigm, Hyperbolic Hypergraph Convolutional Network (HHGCN), which integrates multi-space modelling and higher-order relational learning to more precisely quantify social influence. Specifically, we design a method for learning topological features and capturing group-level environmental characteristics, then transform multi-type social networks that include follow and retweet relationships into a unified hypergraph. Leveraging hyperbolic space and hyperbolic hypergraph operations, our model extracts latent predictive signals. Experiments on eight social network datasets demonstrate that HHGCN outperforms mainstream GNNs on ACC, AUC, and F1, validating the advantages of hyperbolic geometry for hierarchical structures and hypergraphs for higher-order interactions.

Non-intrusive Quality of Service (QoS) monitoring is a fundamental capability for proactive Service Level Agreement (SLA) management and resource optimization in multi-tenant public clouds, yet state-of-the-art models face a dual bottleneck in production deployments due to high-dimensional feature overhead and the prohibitive memory footprint of deep-stacked architectures. To address these challenges, we propose SG-LR QoS, a Sparse-Gating and Layered Low-Rank based lightweight QoS monitoring framework that co-designs data acquisition and model inference under a unified online resource budget. The framework first incorporates AnnealGate, an automated feature selection module representing the sparse-gating component, which employs differentiable $L_{0}$ regularization to learn a structured sparse input subset and prune redundant features. Building upon this, the layered low-rank component is realized through TrunkShare-LoRABank (TS-LB), a parameter-efficient mechanism centered on a shared trunk encoder layer that iteratively reuses parameters across multiple steps while injecting incremental representations via a bank of low-rank adapters. To ensure robust and stable QoS degradation predictions, this recursive process is seamlessly integrated with a Soft Routing Fusion strategy that dynamically aggregates intermediate representations from each step. Experimental evaluations on a large-scale production dataset involving 13 representative cloud applications demonstrate that SG-LR QoS achieves high-fidelity performance with a Mean Absolute Error (MAE) within a 5% margin. Compared to the state-of-the-art model, our method achieves a 40.85% feature sparsity rate while delivering a 79.2% reduction in parameter count, a 45.8% decrease in memory footprint, and a $\mathbf{2. 7 5} \times$ speedup in inference latency.

Modern AI serving increasingly relies on NPUs for conventional inference and large language model serving. However, current NPU deployments commonly expose physical devices directly to applications, which limits runtime control over scheduling and makes it difficult to adapt execution to phase-level workload behavior. This limitation is particularly evident in LLM serving, where the prefill phase is compute-intensive while the decode phase is often constrained by memory bandwidth and KV-cache accesses. Static prefill-decode (PD) disaggregation reduces phase interference, but can introduce resource imbalance and unnecessary data movement. We present FlexNPU, a transparent user-space virtualization layer for Ascend NPUs. FlexNPU interposes on AscendCL APIs and routes NPU operations through per-device daemons, decoupling unmodified from physical NPU devices without modifying model code, AI frameworks, or NPU drivers. This runtime boundary allows FlexNPU to virtualize NPU objects, control operator dispatch, and support phase-aware scheduling for LLM serving. In particular, FlexNPU enables dynamic PD co-location, which adapts scheduling between prefill and decode according to their complementary resource characteristics. We implement FlexNPU on Huawei Ascend NPUs and evaluate it with typical LLM workloads. Compared with direct NPU passthrough, FlexNPU introduces no measurable inference overhead and slightly improves throughput in some scenarios. On a 384-card Ascend 910C deployment of DeepSeek-R1, FlexNPU improves throughput over static PD disaggregation by 5.15

Recent advances in temporal graph research have redefined traditional static graph concepts such as triangles, motifs, and k-cores. Inspired by this, we introduce a novel (k,δ)-truss for temporal graphs, requiring triangles to exist within sufficiently short time windows. The (k,δ)-truss ensures both static and temporal cohesion, while the original k-truss is a special case when δ= ∞. To address (k,δ)-truss queries, we propose index-free and index-based approaches. Utilizing the dual containment relation of (k,δ)-trusses, our indexes losslessly compress all (k,δ)-trusses into map or tree structures, significantly reducing space while enabling optimal-time retrieval. To scale to large temporal graphs, we develop two index construction algorithms based on truss decomposition and truss maintenance, respectively, which substantially reduce redundant computations. Moreover, we present techniques for the dynamic maintenance of the proposed indexes. The experimental results demonstrate that index-based approaches process queries in interactive time and outperform the index-free approach by 2∼4 orders of magnitude, while the indexes achieve compression ratios of up to 10^-4 and can be updated efficiently without rebuilding from scratch.

Repurchase prediction is critical for e-commerce platforms to optimize customer retention, particularly during major promotional events like Double 11. Traditional machine learning models relying on handcrafted features often struggle to capture complex sequential behavioral patterns and dynamic temporal dependencies. To address this issue, we propose a novel two-stage ensemble framework that seamlessly integrates a lightweight Transformer-based sequence model with Gradient Boosting Decision Trees (GBDTs). In the first stage, we construct a comprehensive multi-granularity feature engineering module that extracts user-merchant interaction statistics, item-level aggregations, and multi-window temporal trends, enhanced by Bayesian smoothing. Concurrently, a tailored, lightweight Transformer extracts dense behavioral representations. Specifically, it employs multi-head self-attention and a Last-K pooling strategy to capture bursty, immediate purchase intents while filtering out long-tail browsing noise without relying on user-ID embeddings. In the second stage, a Stacking ensemble strategy combines the predictive probabilities of the Transformer and multiple GBDT models (i.e., LightGBM, XGBoost, CatBoost). We apply a Logit transformation to normalize these probability distributions before utilizing a Logistic Regression meta-learner for final prediction. Experiments on a real-world e-commerce dataset demonstrate that our decoupled framework achieves superior performance compared to other state-of-the-art methods. This study highlights the effectiveness of treating deep sequence models as dedicated feature extractors alongside comprehensive statistical learning.