Large language models now serve enterprise systems, where they plan work, execute handoffs, review artifacts, and improve prior outputs rather than only answer isolated chat requests. Most safety benchmarks test single prompts or jailbreak dialogues and miss how unsafe intent can flow through workflow history, inherited subtasks, and intermediate artifacts. We present JailFlowEval, a workflow-safety benchmark that turns harmful intents into stage-specific records for four common workflow contexts: plan, exec, review, and improve. Its two-phase pipeline builds workflow state and rewrites refusal-triggering text while keeping the unsafe goal; the released high-quality subset contains 744 workflow evaluation records. Tests on 10 models show that workflow-aware subtasks reveal unsafe continuation risks missed by prompt-level evaluation. The main risk context varies by model and metric. This setup maps risk by stage and helps place safeguards where unsafe goals persist as work moves through later service steps after the initial user request is processed.
Real-world time series data is often highly non-stationary, typically exhibiting low-frequency dominated long-term trends and periodic structures alongside high-frequency short-term fluctuations and noise. These distinct frequency components are highly coupled within the time domain. Existing time-domain decomposition methods typically use fixed-size smoothing windows for trend extraction. This approach often blurs information or attenuates local dynamic features, making precise frequency isolation difficult. Furthermore, it is challenging for a single model to simultaneously balance long-range dependency modeling and local detail modeling, thereby limiting the accuracy of long-term forecasting. To address these issues, this paper proposes a dual-branch time series forecasting approach based on frequency decomposition. First, Reversible Instance Normalization (RevIN) is applied to mitigate the non-stationary statistical distribution shifts in the data. Subsequently, Fast Fourier Transform (FFT) is utilized to decompose the sequence in the frequency domain before transforming it back into the time domain. During the forecasting phase, a dual-branch architecture is designed: a Low-frequency Patch-based Forecasting Block (PFB) processes the low-frequency components, while a High-frequency Convolutional Forecasting Block (CFB) handles the high-frequency components. This design not only captures long-term trends and periodic structures via the patch mechanism and self-attention, but also fully leverages dilated causal convolutions to characterize local abrupt changes and short-term volatility. Finally, the prediction results from both branches are fused and denormalized to yield the final output sequence. Experimental results on seven benchmark datasets demonstrate that the proposed method achieves significant performance improvements in long-term time series forecasting tasks.
Environmental, Social, and Governance (ESG) considerations have become increasingly important in modern financial management; however, many existing ESG-oriented analytical models emphasize predictive accuracy while offering limited interpretability and weak support for actionable financial decision-making. An ESG-aware intelligent financial management framework operating as a sequential pipeline is proposed in this study. First, TabNet employs a sparse attention mechanism for instance-wise feature selection of influential ESG and financial indicators, ensuring interpretability and transparent identification of key sustainability drivers. The selected features are then passed to a deep neural network–based regression model to capture complex nonlinear relationships between ESG attributes and financial performance outcomes and generate accurate predictions. Finally, a constrained decision optimisation layer transforms the predicted outputs into actionable capital allocation and investment prioritisation strategies under sustainability and risk constraints. Experimental evaluation conducted on a comprehensive ESG and financial performance dataset demonstrates that the proposed framework achieves improved predictive accuracy compared with conventional regression and machine learning baselines, with a root mean squared error of 4.8617. The results further highlight the varying influence of environmental, social, and governance factors on financial outcomes and confirm the interpretability benefits of sparse attention mechanisms. Overall, the proposed framework provides an effective and transparent decision-support tool for ESG-driven financial management, bridging the gap between predictive modelling and practical, sustainability-oriented financial decision optimisation.
The rapid expansion of digital trade platforms has led to increased transaction volumes, necessitating advanced fraud detection mechanisms to mitigate financial risks and operational threats. Traditional anomaly detection methods often struggle to capture complex and dynamic patterns within highly interconnected trade networks. This study proposes a novel framework based on Graph Neural Networks (GNNs), incorporating Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT), to identify abnormal transactions. The approach utilizes the GlobalTradeSettleNet dataset, applying pre-processing steps such as data cleaning, normalization, and feature transformation. Key features including transaction volume, frequency, node connectivity, and temporal behaviour are modelled within a graph structure. Through neighbourhood aggregation, the model captures intricate relationships among entities and generates anomaly scores. Experimental results demonstrate strong performance, achieving 97.23
This paper introduces a traffic signal control system using deep reinforcement learning to solve the congestion problems at signalized intersections under dynamic traffic conditions. The proposed framework is simulated with the help of MATLAB-SUMO co-simulation framework, the traffic signal control is modeled as a Markov Decision Process (MDP). State space includes traffic density, the queue length, vehicles waiting time, and the actual signal phase whereas the action space comprises of the possible selections of the signal phase. A Deep Q-Network (DQN) is utilized to estimate the optimal state-action value function so that green times can be dynamically allocated based on the changing traffic demand. Multi-objective reward functionality is based on the combined minimization of vehicle delay, queue length, and waiting time and maximization of traffic throughput. Experience replay and target network updates are used to stabilize the learning process. Simulation experiments are conducted in low, medium, and high traffic demand conditions to compare the work of the suggested framework with fixed-time control, actuated control, and classical tabular Q-learning methods. Experimental results demonstrate that the proposed framework reduces average delay by up to 33.9
Enterprise service systems increasingly require autonomous service composition and dynamic reconfiguration to automate complex workflows across organizational boundaries. LLM-based AI agents are being deployed to orchestrate specialized services for tasks ranging from supply chain coordination to business process optimization. A critical challenge in this autonomous service engineering is skill management: how can agents automatically discover, compose, and reconfigure service invocations in response to varying business contexts? Existing approaches treat skills as independent entities activated through semantic matching, failing to capture the sequential dependencies and context-dependent relationships governing effective service composition. We identify two critical challenges. First, skills exhibit long-range dependencies where one service invocation enables another several steps later, defying standard Markovian modeling. Second, skill relationships vary with operational context: optimal sequences differ between production validation and rapid prototyping. Drawing from computational neuroscience on hippocampal cognitive maps and episodic memory, we propose Skill Cognitive Maps for autonomous service planning. Our method learns a task-agnostic cognitive map capturing general skill dependencies through episodic trajectory embeddings, then adapts to specific contexts via context-gated planning. Our dual memory system enables episodic matrices to capture trajectory-level dependencies for generalization across service workflows, while task-specific patterns enable exact sequence retrieval for recurring business processes. Unlike methods requiring massive user data, our centralized approach constructs a single cognitive map that generalizes across tasks while adapting to context through efficient online updates, supporting continuous service lifecycle management. We provide theoretical convergence guarantees and analyze construction and retrieval complexity. Experiments on three datasets demonstrate superior performance over Context-Aware Q-Learning, LLM-Chain, ToolFormer, and Semantic-Tool, validating our dual memory architecture for autonomous service composition in enterprise value chains.
Crimes and accidents can have significant economic effects on not only people but also entire towns and countries. Predictive policing is a strategy that uses data analytics techniques, machine learning algorithms, and statistical models to predict the location and time of the likely events and individuals who might be involved in unlawful actions. Predictive policing uses data analytics, machine learning algorithms, and statistical modelling to support proactive policing and the allocation of operational resources. This paper examines the emergence of algorithmic bias in such systems and evaluates practical methods for its detection and mitigation. Using the Chicago Crime Dataset, the study applies three complementary techniques—data re-weighting, counterfactual analysis, and algorithm auditing—to identify both data-driven and model-driven sources of inequity. The results show that conventional predictive models disproportionately classify minority and historically over-policed neighbourhoods as high-risk, with risk estimates driven primarily by variables reflecting enforcement intensity rather than underlying crime incidence. Bias-controlled models produce materially different hotspot predictions and reduce measures of disparate impact, indicating that historical policing practices significantly shape risk forecasts. The findings highlight the need for transparency, fairness-aware modelling, and systematic auditing in the development and deployment of predictive policing systems. The paper presents an analysis of various techniques for diagnosing and mitigating bias in algorithmic crime prediction, contributing to a more accountable and evidence-based use of AI in law enforcement contexts.
Accurate and efficient detection of software defects in complex and high-dimensional datasets is one of the biggest challenges in software development. For this purpose, the study proposes the AGT-GA-LSTM framework, which integrates gate-adjustable LSTM and the Artificial Gorilla Troops optimizer. Defect detection and management are essential at these stages to guarantee the software’s dependability and quality. This research focuses on enhancing defect detection through the integration of deep learning (DL) techniques, by proposing a novel model, the Artificial Gorilla Troops optimizer-tuned Gate Adjustable Long Short-Term Memory (AGT-GA-LSTM). Using DL to forecast software faults, this research tries to create a strong framework for defect detection. Publicly available software module datasets are utilized. Before feeding the data into the model, Z-score normalization is applied during the preprocessing phase to enhance the quality of the data. Feature extraction is performed using Kernel Principal Component Analysis (Kernel-PCA) to reduce dimensionality while retaining essential information, improving the model’s performance, and then feeding the processed data into the AGT-GA-LSTM model. This model combines the Artificial Gorilla Troops optimizer, which enhances the LSTM’s performance, with gate mechanisms to adaptively filter important features for prediction tasks. The results show that the AGT-GA-LSTM performs better than traditional machine learning models, with a 92.4
AI chatbots are increasingly embedded in e-commerce platforms to provide automated and real-time customer service in business-to-consumer settings. Grounded in expectation confirmation theory, this study examined how online shoppers constructed expectations, evaluated their experiences with AI chatbot services, and perceived effective collaboration between AI chatbots and human agents. A mixed-methods research design was adopted, beginning with a repeated-measures experimental approach. Data were collected from 30 participants, yielding 120 data points for quantitative analysis. This was complemented by World Café focus group discussions to further explore customer perspectives on collaboration between AI chatbots and human agents. The findings indicated that AI chatbot performance in handling complaints and problems, managing customer expectations and satisfaction, and influencing customer churn intention significantly affected customers’ electronic word-of-mouth (eWOM) behaviors. Satisfied users were more likely to share positive online feedback, whereas dissatisfaction increased negative reviews and customer churn intentions. Accordingly, improving service quality through an AI chatbot is critical for an e-commerce platform to facilitate customer retention and a firm’s online reputation. Particularly, training empathetic chatbots and creating user-friendly interfaces help mitigate customer-perceived impersonality of robots and foster stronger customer engagement. When AI chatbot service failures occur, well-designed service recovery mechanisms, including timely referral to human agents, can effectively reduce customer dissatisfaction and negative eWOM behaviors.
Artificial Intelligence (AI) has significantly revolutionized customer service across all sectors, introducing disruptive and transformative technologies to enhance efficiency, responsiveness, and overall user experience. This article investigates AI-based customer service, thoroughly exploring current applications and offering insights into potential future directions for development. This work conducts an in-depth multimethod research analysis of existing AI applications in customer service, emphasizing the role of Natural Language Processing (NLP), Machine Learning (ML) and vision-based systems. The integration of predictive analytics is examined as a key component, enabling the interpretation of customer needs, and predicting potential issues, thereby facilitating proactive resolution. Moreover, the article investigates the impact of AI-based personalization on customer interactions, providing a comprehensive discussion on how recommendation engines and personalized content delivery contribute to heightened customer satisfaction and loyalty. The study also confronts challenges associated with AI implementation in customer service, addressing concerns related to data privacy, ethical considerations, and continuous monitoring and improvement. The findings of this study contribute to the ongoing scientific discourse on the integration of AI in customer service, offering insights for researchers, practitioners, and stakeholders in the service industry seeking to fully leverage the potential of AI-driven customer interactions. Anticipating the future, this research delineates prospective paths for AI in customer service, underscoring the potential of advanced conversational agents, virtual assistants, and emotionally intelligent AI systems.
The dynamic and distributed nature of cloud-native systems makes it challenging to ensure high availability, which poses serious operational challenges. Standard Kubernetes readiness and liveness probes and other health monitoring systems rely on binary, simplistic checks that miss important but subtle "gray failures" like memory pressure, resource contention, and downstream service degradation. A new framework for autonomous anomaly detection and real-time remediation in cloud-native systems, eBPF-Shield, is presented in this paper. By utilizing the extended Berkeley Packet Filter (eBPF), eBPF-Shield creates a closed-loop control system that functions as a hybrid architecture across user-space and the Linux kernel. By utilizing an adaptive baselining model to learn the typical behavior of every container from a comprehensive, multi-modal stream of telemetry covering network, CPU, memory, and I/O subsystems, it surpasses static, pre-configured thresholds. In order to stop cascading failures, the system immediately implements proactive remediation, like graduated load shedding, in the kernel’s data path based on a continuously calculated health score. We put eBPF-Shield into practice and tested it using a realistic microservices benchmark by putting it through a number of fault injection scenarios. In comparison to state-of-the-art baselines, our evaluation shows that eBPF-Shield reduces the time-to-detect for latency-based anomalies by up to 90
Service-oriented computing enables service reuse across different contexts, with applications heavily relying on third-party lightweight HTTP services. Testing such applications, however, presents significant challenges when actual services are inaccessible. Service virtualisation provides a technique to overcome dependency issues by simulating service behaviour through synthetic responses. This paper investigates the suitability of Description Logic learning as a symbolic machine learning technique for automatically constructing HTTP service mock skeletons from network traffic recordings. Experiments demonstrate strong predictive performance with human-readable logical expressions for key response attributes including status codes and headers, capturing both protocol semantics and temporal dependencies on prior transaction sequences. The resulting mock skeletons facilitate comprehension of the logical reasoning behind response properties and can be customised to create mocks that generate responses suitable for testing.
Availability and reliability are essential prerequisites for web service-based systems to function seamlessly. However, dynamic operating environments expose web services to potential challenges like failures, unavailability, and fluctuating quality of service. Such failures in any component service can degrade the performance of an entire composite service. This paper presents a proactive web service selection approach that handles runtime failures by identifying optimal replacement candidates during the initial composition stage. The proposed method combines multi-criteria QoS evaluation using the PROMETHEE Plus method, checks for semantic similarity through a determinant-based IOPE (Input, Output, Precondition, Effect) matchmaking mechanism, and uses a genetic algorithm to optimize service composition plans. First, QoS attributes of candidate services and user requirements are normalized to a common scale. They are then evaluated using PROMETHEE Plus, which produces a prioritized ranking based on Net Outranking Flow scores. At the same time, functional compatibility between the requested and candidate services is assessed through IOPE-based semantic matchmaking. This involves constructing weighted bipartite graphs and solving them using a determinant-based maximum-matching approach to compute similarity scores and ranks. Next, these QoS rankings and semantic similarity measures are combined to estimate service replaceability. Finally, a genetic algorithm evolves candidate compositions iteratively. Each chromosome represents a service combination. A hybrid fitness function that combines QoS attributes and replaceability scores guides selection toward optimal, resilient composition plans. Experimental results demonstrate that the method achieves superior selection quality compared to existing techniques, with average improvements of 9.7
In modern software systems, Web services are the primary means of provisioning remote resources. The service ecosystem exhibits significant duplication, with multiple services fulfilling, in essence, the same functional requirements. Such services often can be used interchangeably, so developers face the challenge of choosing the most suitable service for the task at hand. However, extant approaches for selecting services focus solely on the service’s performance characteristics (so-called QoS, whose properties include latency, reliability, availability, etc.). These approaches often neglect important utility characteristics (e.g., data accuracy, correctness, coverage, etc.). As a consequence, selected services may exhibit high performance, while delivering information that is either inaccurate or outdated. This article addresses this problem by introducing Quality of Information (QoI), a quality metric that measures data-related service performance. To be able to measure QoI accurately and effectively, we classify aspects of service QoI (e.g., data freshness) that can be measured automatically by comparing the service outputs and those (e.g., accuracy) that require manual effort for labeling the ground truth of a given test input. To accurately measure QoI without the costly manual effort, we formulate the input selection problem as selecting a small set of test inputs to most accurately approximate the QoI obtained by using a large input set. Inspired by input sample selection methods for testing machine learning (ML) algorithms, we have adapted these methods to evaluate their applicability on two datasets of service invocation results, having observed a noteworthy performance variance between ML testing and web service testing. Having identified insights and challenges in measuring QoI, this research highlights the need for further investigation.
In service-oriented computing environments, intelligent service ecosystems are characterized by dynamic service composition, evolving user preferences, and heterogeneous QoS constraints. Accurate modeling of user–service interaction sequences is therefore essential for adaptive and QoS-aware recommendations. However, existing sequential recommendation models typically rely on a single architectural paradigm, limiting their ability to capture multi-granularity behavioral dynamics. To address these challenges, we propose MageRec: a Parallel Mamba–Attention Gated Network for Sequential Recommendation. MageRec adopts a parallel architecture in which the Mamba and attention branches jointly model global contextual dependencies in user interaction sequences. Meanwhile, a local convolution module is introduced to extract local behavior patterns within short interaction neighborhoods. A context bias network further enables adaptive fusion of these heterogeneous signals under evolving service conditions. Extensive experiments on multiple public datasets demonstrate that MageRec consistently outperforms existing baselines. Our code is made publicly available on https://github.com/cccsama/MageRec .
The everyday rapid growth of the Internet of Things(IoT) across many domains like smart home, healthcare, agriculture, and industries has introduced increasingly complex security challenges. This paper presents a comprehensive survey of application-specific IoT security frameworks, highlighting their challenges and strategies, and strengths or limitations. The paper further surveys existing standards, regulations, as well as common attack vectors, including malware, DDoS, firmware tampering, and side channel analysis. In this review, the literature is organized into three stages:(i) IoT security frameworks covering layered architectures, general-purpose, AIML, or application-specific, (ii) security requirements and solutions across major IoT application domains; and (iii) Different IoT security mechanisms spanning hardware, network, and data layers. In addition to surveying and synthesis, this study introduces a novel LLM-based assistant architecture for domain-specific IoT security framework generation. The proposed expert assistant system leverages a fine-tuned open-source LLM or a custom-trained LLM to interpret high-level user intent, including IoT application, IoT device, Hardware, software, special concern objectives, and translates them into transparent, explainable, and actionable security framework suggestions. This review closes the gap between conventional static security models and one-size-fits-all security frameworks and the growing demand for customized, practical, and domain-specific IoT security design, providing both a critical synthesis of current research and a forward-looking AI-assisted framework for intelligent IoT security planning.
Explainable prediction from complex data like text and time series is a critical challenge for intelligent service-oriented applications. Traditional models often lack transparency, hindering trust in service contexts where understanding data nuances is key. Dynamic pricing in online rental services like Airbnb exemplifies this, where explainable price forecasts based on factors such as guest reviews are vital for service optimization. Large Language Models (LLMs) show promise for explainability but face challenges with noisy data interpretation and costly expert annotation for fine-tuning in these service-specific tasks. To address these issues, we introduce the Guided Self-Reflection framework. Our key insight is enabling an LLM to autonomously generate high-quality “Golden data” from public service information (e.g., Airbnb listings and reviews) via a structured, iterative self-reflection process. This reflection is meticulously guided by content-specific criteria: Evidence Relevance and Faithfulness, Logical Consistency and Coherence, and Domain Commonsense Alignment (pertinent to Airbnb price dynamics). This process teaches the LLM to discern and articulate the interplay between textual feedback and price movements. The LLM agent then leverages this autonomously generated data, including positive and negative reflective examples, to self-fine-tune its predictive and explanatory capabilities using Proximal Policy Optimization (PPO), thereby eliminating the need for costly manual annotation. Applying this framework, we fine-tune a specialized LLM for explainable predictions of Airbnb rental price movements, conditioned on historical prices and contemporary guest reviews. Our methodology demonstrates potential to significantly outperform traditional deep learning techniques and other LLM applications in both predictive accuracy and the Matthews Correlation Coefficient (MCC). Critically, it offers a pathway to generating transparent and trustworthy explanations for price dynamics, enhancing the intelligence and user-centricity of service-oriented pricing applications and contributing to more robust, interpretable AI solutions in the service industry.
In edge computing environments with multiple edge servers, load imbalance is common due to varying user densities. High-density regions often lead to overloaded servers, while low-density areas result in resource underutilization, reducing overall system efficiency. To address this, task migration from high-load to low-load servers is a feasible solution. However, as edge servers are often self-interested, they may not willingly offer computing services for free. To incentivize cooperation, we propose a task migration strategy based on a double auction mechanism, termed Double Auction Task Migration Strategy (DATMS). In this strategy, users offload tasks to their nearest servers. High-load servers (sellers) determine selling prices and task quantities to migrate, while low-load servers (buyers) decide their bid prices for the migrated tasks. The double auction mechanism then matches sellers with buyers to maximize mutual benefits and achieve system-wide load balancing. The sellers’ decision-making process is modeled as a Markov Decision Process (MDP) and solved using the Deep Q-Network (DQN) algorithm. For buyers, who operate under partial observability and continuous action spaces, we model the problem as a Partially Observable Markov Decision Process (POMDP) and adopt the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm. The integration of DQN, MADDPG, and the double auction mechanism forms the complete DATMS strategy. Extensive experiments demonstrate that DATMS significantly outperforms four benchmark strategies in terms of long-term load balancing across the entire system.
To address the issues of weak relational semantic perception, deep-layer information attenuation, and insufficient structural utilization in existing knowledge graph-based recommendation models, this paper proposes a Multi-layer Knowledge Graph-based Recommendation model with Relational Attention (MKGRA).The model consists of three core modules: an entity-relation dual attention mechanism to enhance semantic perception of heterogeneous relations; a multi-layer residual propagation structure to mitigate deep-layer information attenuation; and a graph neural network-based semantic fusion module to improve the utilization of graph structural information. Experiments comparing with 19 baseline models on three datasets (Last-FM, Book-Crossing, and Yelp2018) show that MKGRA significantly outperforms mainstream baseline models on multiple metrics. Its AUC reaches 0.848, 0.754, and 0.881 respectively, while its ACC is 0.771, 0.690, and 0.839.The experimental results verify the effectiveness of each module, providing a more accurate method for semantic modeling and structural mining in knowledge graph-based recommendations.
The issue of ensuring data consistency across different domains in cloud-based distributed storage systems is a major problem, being mainly due to the maintenance of synchronization and the reliability of the distribution of data across various domains. In this research, these issues are tackled by the introduction of the Addax Optimization Algorithm (AOA), which is a new method that synergistically combines optimization techniques with K-means clustering to support data consistency. The overall goal is to attain simultaneous reduction of consistency latency and maximization of packet delivery ratio (PDR) and data integrity in multi-domain settings. In the course of the experimental evaluation, the new technique showed better performance than the conventional approaches such as Particle Swarm Optimization (PSO) and Genetic Algorithm (GA) on the important parameters of latency, PDR, and data integrity, thus proving its suitability for real-world cloud-based distributed systems. The findings imply that AOA is a good choice for facilitating cross-domain data consistency with remarkable advantages in terms of scalability and reliability even in diverse network conditions.
