With the rapid growth of mobile internet services, the number of applications (apps) has increased dramatically, leading to severe information overload in app markets. Efficiently identifying personalized and relevant apps for users has thus become a critical issue in intelligent recommendation systems. Existing approaches, including sequential modeling, graph neural network-based intearaction mining, and large language model-driven semantic modeling, have achieved notable progress. However, challenges such as the neglect of group-level collaborative signals, the embedding redundancy and inefficient feature integration, and the disjoint modeling of short-term intent and long-term preferences still hinder recommendation performance. To address these issues, this paper proposes a multi-view fusion framework integrating hypergraph structures and llm-enhanced semantics for mobile app recommendation (MHLAR). The framework first establishes a stable foundation via a behavioral pre-training module, where an SR-GNN encoder learns short-term transition patterns from session sequences. Building upon these robust behavioral representations, MHLAR progressively integrates multi-dimensional long-term preference signals through gated fusion and contrastive learning, facilitating a collaborative multi-view architecture for personalized app recommendation. Experiments on the public MobileRec dataset demonstrate that MHLAR consistently outperforms a wide range of state-of-the-art baselines across all evaluation metrics, yielding substantial gains in both recall and ndcg. Ablation studies validate the effectiveness of each core component, and hyperparameter sensitivity analyses provide deployment guidance.

The innovative smart contract mechanism underlying Ethereum has accelerated the growth of its ecosystem. However, the financial transaction characteristics of Ethereum have also facilitated illegal activities, such as phishing scams. These scams pose a serious threat to the security of the ecosystem. Current phishing scam detection methods on Ethereum rely on transaction network graph structures that struggle to capture the complex behavioral patterns of accounts, resulting in low detection efficiency. To address this challenge, this paper proposes a multi-view hypergraph fusion (MVHGF) framework designed to model the distinctive characteristics of Ethereum transaction networks. We design four hypergraph generation strategies from both local (account transaction behavior and transaction motif substructures) and global (transaction network topology and account attributes) perspectives, thereby modeling complex behavioral patterns. Subsequently, hypergraph convolutional neural networks are employed to learn a representation vector for each hypergraph. An adaptive integration mechanism is then used to fuse the multi-view representations, thereby integrating key features from different perspectives. Experimental results show that MVHGF significantly outperforms the single-view graph learning baseline TSGN, improving the AUC from 92.36% to 96.42% and the F1-score from 91.12% to 96.23%, demonstrating excellent detection accuracy and robustness. This work provides a novel approach to detecting phishing scam accounts on Ethereum by effectively fusing multi-view information.

Cloud computing optimizes service quality and resource efficiency via centralized hardware and computational resources. However, the predominantly centralized deployment and operation of cloud data centers increase the physical distance to end-users, leading to degraded service quality. Edge computing addresses this by offloading data processing and analysis tasks directly to devices at the network edge, reducing reliance on backhaul transmission and thus offering a more responsive solution for latency-sensitive applications. Nevertheless, ensuring that applications meet predefined Service Level Agreement (SLA) in resource-constrained edge environments remains challenging. To tackle these issues, this paper investigates elastic scaling strategies in cloud-edge collaborative settings. We propose an attention-enhanced bidirectional LSTM model (A-Bi-LSTM) for microservice workload prediction, and design an adaptive elastic scaling system named XScale. This system incorporates a fall-back scaling mechanism when predictions are unreliable and introduces a proactive load forwarding strategy to enhance overall edge node performance. Experimental results show that, compared to existing elastic scaling methods, XScale reduces SLA violations by 82.3%, increases average resource utilization by 17.4%, decreases average response time by 21.1%, and improves overall edge node performance by 36.3%.

Recommender systems are pivotal in enhancing user experience on a variety of platforms by providing personalized suggestions based on user behaviors, historical interactions, and item attributes. Recently, recommendation methods combining graph neural networks with contrastive learning have gained attention for their superior performance. However, despite their success, they still face challenges in capturing structural information in the user-item interaction graph and are sensitive to noisy user behavior and the quality of contrast views. To address these limitations, this paper proposes a novel multi-view hypergraph contrastive learning method for item recommendation (VGCL). Specifically, the method involves constructing user-item interaction view, user-user collaborative view, and item-item collaborative view by leveraging user information, item information, and user-item interactions. It then learns user and item feature vectors by incorporating semantic and higher-order correlation information through graph and hypergraph neural network techniques. The method performs contrastive learning on the user and item feature vectors to enhance their quality. Finally, it predicts user preference scores for items using multi-layer perceptron and recommends the most suitable items to users. Extensive experiments on various real datasets show that VGCL outperforms existing state-of-the-art methods.

A non-fungible token (NFT) is a unique digital asset stored on the blockchain. With growing interests from investors and collectors, understanding the factors influencing the NFT market is crucial. This study investigates the influence of Twitter social media and trading patterns on transaction volume in the CryptoPunks NFT market. We focus on the influencing factors such as trading behavior, Twitter sentiment, and Twitter features. Based on them, we employ machine learning to predict the CryptoPunks transaction volume. The results reveal a strong correlation between these features and transaction volumes. The proposed PSO-Voting model, which takes all features into consideration, can improve the predictive accuracy significantly, achieving an impressive model accuracy (R2) exceeding 96

With the growing number of services, developers struggle to efficiently discover and compose suitable ones for Mashup development, underscoring the need for intelligent service recommendation. Although recent graph- and hypergraph-based methods with contrastive learning show promise, they still face challenges such as degraded feature representations, fuzzy high-order association modeling, hyperedge sparsity, biased negative sampling, and entangled feature couplings. To address these issues, we propose a Multi-view heterogeneous Hypergraph augmented self-Gating Contrastive Fusion framework (MHGCF) for service recommendation. MHGCF constructs heterogeneous hypergraphs from interaction, semantic, and category views to capture diverse mashup–service relations. A self-gating mechanism refines neighbor aggregation to suppress noise and retain discriminative features, while a cross-view attention module enables fine-grained fusion and disentanglement. Furthermore, hypergraph perturbation creates structural views for contrastive learning, enhancing representation consistency and alleviating sparsity. Experiments on a real-world dataset show that MHGCF outperforms state-of-the-art methods, demonstrating its effectiveness and superiority.

The rapidly expanding Ethereum ecosystem has driven the flourishing of decentralized applications, but has also brought increasingly severe security risks. Ponzi scheme, in particular, pose a grave threat to platform security and user assets by luring investors with promises of high returns. The current detection methods generally suffer from limitations such as insufficient feature extraction, reliance on a single information source, and poor robustness. To address these challenges, this paper proposes a novel Multi-View Multi-Modal Fusion Framework with Large Language Models for Ponzi scheme detection on Ethereum, named MF2LLM. We first model the contract opcode sequence as an opcode chain graph and design a Time-Stamped Graph Encoder (TS-GE) to capture local temporal dependencies and execution flow relationships between opcodes. Concurrently, we construct an opcode semantic hypergraph based on semantic categories and design a Semantic-Weighted Hypergraph Encoder (SW-HGE) to model higher-order co-occurrence patterns and global associative features. Furthermore, we propose the Opcode Sequence Lightweighting (OSL) method, which significantly compresses the length of opcode sequences while preserving core control logic and semantic information. This provides high-quality structured input for information fusion. To this end, we perform multi-modal instruction fusion on multi-source heterogeneous features and employ LoRA to fine-tune LLMs. This enables the model to achieve cross-modal semantic reasoning and behavioural pattern recognition. Through extensive experimental validation on real-world datasets, MF2LLM demonstrates stable and superior detection performance even under conditions of highly imbalanced sample distributions. Compared to existing state-of-the-art approaches, our method outperforms across all metrics, achieving an ACC of 99.43%, Precision of 96.57%, Recall of 97.06%, and an F1-score of 96.81%. The efficiency and practical value of MF2LLM in detecting Ponzi schemes on Ethereum contribute to enhanced security for the decentralized application ecosystem. The codes are publicly available on Github: https://github.com/yemisua/MF2LLM.

The rapid development of the Ethereum platform has been accompanied by increasingly severe security challenges, among which phishing attacks have become one of the most prevalent attack vectors. Attackers exploit the anonymity of blockchain accounts and employ social engineering techniques to induce users to disclose private keys or grant authorizations to malicious smart contracts, resulting in substantial economic losses. Current detection methods struggle to simultaneously capture dynamic temporal patterns, higher-order interactions, and long-range dependencies in transaction networks. To address these challenges, this study proposes an approach that integrates evolving Hypergraph Sequence with Large Language Models for detecting phishing scam accounts on Ethereum, named HyperSeq-LLM. Specifically, we employ a sliding window to construct temporal transaction sequences that capture dynamic transaction patterns, and we develop a hypergraph generation method based on feature clustering to model multi-party higher-order interactions within these sequences. Additionally, we integrate hyperedge-size-aware weighting into Hypergraph Convolutional Networks (AwareHGCN) to enhance the representation of groups. Finally, we employ quantized low-rank adaptation techniques to fine-tune large language models to capture the evolution of account transaction patterns within temporal hypergraphs. Experiments on the Ether@P1:N1 and Ether@P1:N10 datasets demonstrate that HyperSeq-LLM consistently surpasses state-of-the-art baselines, exhibiting enhanced detection accuracy and early-warning capabilities. These findings underscore the efficacy of integrating dynamic modeling with multi-relational interactions and highlight HyperSeq-LLM’s scalability and practical potential for securing the Ethereum ecosystem. The codes are publicly available on Github: https://github.com/yemisua/HyperSeq-LLM

The rapid development of blockchain technology has led to the widespread use of smart contracts for the development of blockchain applications. The vast number of smart contracts presents a considerable challenge for users attempting to locate a specific contract, as the manual process of searching becomes increasingly time-consuming and difficult. To enhance the retrieval of smart contracts, it is imperative to categorize them into distinct groups, thereby facilitating their expedient identification and retrieval. A number of studies have proposed methods for the automated classification of smart contracts, exploiting either the source code, bytecode, or application binary interfaces (ABIs). However, none of the aforementioned studies employ multimodal data. Given that the source code of smart contracts is typically not accessible to the public, we consider both the ABI and bytecode of these contracts and propose a novel classification method for smart contracts based on Hybrid Convolutional Neural Networks, which we have named HCNN-SCC. Specifically, the method employs a convolutional network with an attention module to extract information from the ABI-term matrix, which is derived from the ABI descriptions of smart contracts. Furthermore, a graph convolutional network is employed to extract information from the control flow graphs derived from the bytecode of smart contracts. Finally, the acquired information is aggregated to predict the category of a new smart contract. An experimental study on 3,815 real-world Ethereum smart contracts demonstrates that our proposed method outperforms other baseline techniques.

In mobile edge computing (MEC) environment, effective microservices deployment significantly reduces vendor costs and minimizes application latency. However, existing literatures overlook the impact of dynamic characteristics such as the frequency of user requests and geographical location, and lack in-depth consideration of the types of microservices and their interaction frequencies. To address these issues, we propose TP-MDU, a novel two-stage deployment framework for microservices. This framework is designed to learn users' dynamic behaviors and introduces, for the first time, a minimal deployment unit. Initially, TP-MDU generates minimal deployment units online, tailored to the types of microservices and their interaction frequencies. In the initial deployment phase, aiming for load balancing, it employs a simulated annealing algorithm to achieve a superior deployment plan. During the optimization scheduling phase, it utilizes reinforcement learning algorithms and introduces dynamic information and new optimization objectives. Previous deployment plans serve as the initial state for policy learning, thus facilitating more optimal deployment decisions. This paper evaluates the performance of TP-MDU using a real dataset from Australia's EUA and some related synthetic data. The experimental results indicate that TP-MDU outperforms other representative algorithms in performance.

The latency of cloud-hosted composite applications increases due to extended transmission time from the centralized cloud to end-users, compromising service quality. Typical AI application scenarios like autonomous driving and smart cities demand low network latency. Edge computing addresses this by enabling data collection and analysis in nearby edge data centers, reducing user response time. However, in a complex edge-cloud computing environment, finding the optimal scaling scheme dynamically is crucial due to varying user response times and dynamic scaling costs near edge data centers. This paper proposes a predictive scaling method to adjust microservice container number based on user request fluctuations. Our prediction algorithm, a two-way GRU with an attention mechanism named A-Bi-GRU, aims to minimize scaling jitter. To achieve this, we introduce the concept of an observation window and employ a multi-objective optimization algorithm based on improved NSGA-II, named DP-GA, for microservice scaling across different locations within each window. The solution aims to minimize the average user response time and scaling costs, enabling intelligent dynamic scaling based on location awareness. Experimental results indicate that the proposed A-Bi-GRU forecasting algorithm achieves approximately a 30% improvement in prediction accuracy over traditional linear models such as LR and SVM, and about a 5-10% improvement compared to conventional recurrent neural networks like RNN and LSTM. Furthermore, the proposed DP-GA multi-objective optimization algorithm reduces average response time by roughly 80% and scaling cost by approximately 50%.

The emergence of Ethereum has created new opportunities for decentralized finance. However, the increasing prevalence of illegal activities, such as phishing scams, poses a significant threat to the healthy development of the Ethereum ecosystem. To address this issue, this study proposes an approach that leverages Temporally Evolving Graph Sequences for Detecting Phishing Scams on Ethereum, named GraphSeqGuard. Specifically, we design a novel algorithm to generate dynamically ordered transaction subgraph sequences for each account. To this end, we employ graph convolutional networks as encoders to extract subgraph embedding representations and present local and global subgraph sequence contrastive learning to train the encoder, thereby enhancing its learning capabilities. Additionally, we combine a bidirectional long short-term memory network with a deep state-space module and a multi-head differential attention mechanism to develop a classifier for detecting phishing scam accounts. Evaluated on a real Ethereum dataset, GraphSeqGuard outperforms state-of-the-art approaches in detecting phishing accounts. The proposed approach demonstrates clear advantages in capturing transaction patterns and tracking their dynamic evolution, offering a robust solution for enhancing the security of Ethereum transactions.

The exponential growth of mobile applications (apps) have made it increasingly challenging for users to discover apps that align with their interests. To address this challenge, researchers have drawn inspiration from the success of recommender systems in various domains, such as e-commerce, and developed app recommendation methods. However, existing approaches still face significant limitations, including over-smoothing, noise interference in high-dimensional data, semantic loss, and low-quality positive and negative samples, all of which hinder recommendation performance. To tackle the above limitations, this study proposes HCAppRec, a novel app recommendation approach that leverages user-app interaction history and integrates hypergraph neural networks with contrastive learning. HCAppRec first constructs a couple of hypergraphs by exploring the semantic similarities between users and between apps, derived from the user-app interaction data. Then, by integrating hypergraph neural networks with contrastive learning, HCAppRec can not only capture complex high-order relationships among users and apps but also distinguish subtle differences, enhancing the model's robustness and generalization. Extensive experiments on real-world datasets demonstrated that HCAppRec significantly outperforms state-of-the-art methods in comprehensive recommendation performance.

Blockchain technology is celebrated for its transparency and immutability, revolutionizing trust models. However, its decentralized nature raises concerns about potential inclusion of malicious or illegal content. This study focuses on Ethereum’s blockchain, proposing an algorithm for data identification and restoration. We successfully recovered 175 files, 296 images, and 91,206 texts. Employing FastText for sentiment analysis, we achieved 0.9 accuracy after parameter tuning. Classification revealed 70,189 neutral, 5,208 positive, and 15,810 negative texts, aiding in identifying sensitive or illicit information. Our findings expose benign and harmful content coexisting on Ethereum, including personal data, explicit images, divisive language, and racial discrimination, notably targeting Chinese government officials. This study provides valuable insights for public understanding and regulatory guidance on blockchain technology.

EDGE computing provides consumer electronics with faster and smarter responses, thereby enhancing user experience. Based on container technology, in a cloud-edge collaborative environment, edge servers pull images from image registries in the cloud and deploy them locally to serve various smart devices and consumer electronics. Leveraging image layer sharing can reduce the amount of image transmission, as locally cached layers on edge servers need not be repeatedly pulled. This paper investigates the rational deployment of applications on edge servers under a layer sharing mechanism. Considering a scenario with distributed multiple image registries, the objective is to minimize the total deployment time under the condition of image layer sharing. In this scenario, the application deployment problem can be modeled as a variable-value 0-1 multiple knapsack problem with partial item sharing. To address this improved knapsack problem, two algorithms named greedy minimum value based on layer sharing (GD-LSV) and greedy minimum value density based on layer sharing (GD-LSVD) are proposed. Experimental simulations using real layer dataset have proved the effectiveness of the proposed algorithms. Compared with representative algorithms, as the number of applications increases, the proposed algorithms shorten the total deployment time by up to 51.78% and reduce the total storage consumption by up to 22.79%.

Third-party libraries (TPLs) play a crucial role in software development. Utilizing TPL recommender systems can aid software developers in promptly finding useful TPLs. A number of TPL recommendation approaches have been proposed and among them graph neural network (GNN)-based recommendation is attracting the most attention. However, GNN-based approaches generate node representations through multiple convolutional aggregations, which is prone to introducing noise, resulting in the over-smoothing issue. In addition, due to the high sparsity of labelled data, node representations may be biased in real-world scenarios. To address these issues, this paper presents a TPL recommendation method named Implicit Supervision-assisted Graph Collaborative Filtering (ISGCF). Specifically, it takes the App-TPL interaction relationships as input and employs a popularity-debiased method to generate denoised App and TPL graphs. This reduces the noise introduced during graph convolution and alleviates the over-smoothing issue. It also employs a novel implicitly-supervised loss function to exploit the labelled data to learn enhanced node representations. Extensive experiments on a large-scale real-world dataset demonstrate that ISGCF achieves a significant performance advantage over other state-of-the-art TPL recommendation methods in Recall, NDCG and MAP. The experiments also validate the superiority of ISGCF in mitigating the over-smoothing problem.

Data sharing in Vehicular Social Networks (VSNs) is an essential road service that assists vehicle driving and promotes intelligent transportation applications. In VSNs, vehicles regularly collect and upload valuable data to share with other vehicles. Data encryption can be employed during data uploading and sharing to prevent malicious tampering and privacy disclosure. However, existing data-sharing schemes lack security, have high overhead in obtaining decrypted data, and show low trust in the central authority controlling the entire network. To facilitate data sharing in VSNs, this paper proposes a new scheme using consortium blockchain to realize secure data sharing. Nodes in the blockchain invoke smart contracts and implement the location-based Speculative Byzantine Fault Tolerance (LSBFT) to accomplish data-sharing transactions among vehicles. The scheme not only ensures the security of vehicle information but also protects the privacy of the shared data. Security analysis demonstrates that the proposed scheme can resist attacks and has shown transaction fairness, data confidentiality, non-repudiation, and traceability. Simulation results show that the scheme has higher sharing efficiency and less time to reach a consensus in the data storage process.

Recommendation system plays a remarkable role in solving the problem of information overload on the Internet. Existing research demonstrates that a recommended list enclosed with appropriate explanations can enhance the transparency of the system and encourage users to make decisions. Although existing works have achieved effective results, they still suffer from at least one of the following limitations: the work either does not use sentiment information or review information, does not explicitly incorporate review-level sentiment information into the model, is based on review retrieval, and generates explanations in the form of templates or phrases. To tackle the above limitations, this article proposes a REview-level Sentiment information enhanced multiTask learning approach for Explainable Recommendation (RESTER). Specifically, it first considers the user’s review information and analyzes the sentiment polarity contained in the review. Then, the user/item’s identity feature, review feature, and sentiment information are fused into a multitask learning framework by leveraging the implicit correlation between the rating prediction and explanation generation tasks. Comprehensive experiments on datasets in three different domains have shown that the proposed model is superior to all other baselines in both rating prediction and explanation generation tasks.

Mashup technology enables developers to create new applications more readily by combining existing services. As its popularity grows, research on service recommendation for mashup creation has gained increasing attention. Existing recommendation methods have the following limitations: either they are susceptible to data sparsity problems, or they exhibit over-smoothing when aggregating high-order neighbors, resulting in similar and non-specific node feature representations, or they only focus on bipartite graphs and neglect the rich heterogeneous information in the mashup-service ecosystem. To address these issues, we propose a service recommendation method for mashup creation based on light heterogeneous hypergraph contrastive learning (LHGCL). This method first constructs a heterogeneous hypergraph by combining mashup information, service information, the mashup-service interaction data, and their related attribute information. Then, it designs a light hypergraph neural network to capture the high-order relationships between mashups and services. Next, it applies contrastive learning to enhance the representations of mashups and services. Finally, it utilizes the enhanced feature vectors of mashups and services to predict mashup preferences for services. Comprehensive experiments conducted on the real-world ProgrammableWeb dataset demonstrate the superiority of the proposed method and the effectiveness of its key modules.

With the burgeoning adoption of blockchain technology, cryptocurrencies have surged in popularity, becoming a focal point of global interest. Concurrently, the emergence of cryptocurrency phishing scams poses a significant threat to the financial security of the blockchain ecosystem, inflicting substantial economic damage on platforms and users alike. This study introduces an innovative approach leveraging an attention-augmented Bidirectional Long Short Term Memory Network (BiLSTM), termed BiLSTM4DPS, for the detection of phishing scams within the Ethereum network. We initiate by converting account transaction records into sequences, thereby extracting temporal and latent patterns of transactions. Subsequently, we integrate BiLSTM with multi-head attention mechanisms and masking techniques to construct a robust classification model aimed at identifying fraudulent accounts. Extensive experiments were conducted to assess the efficacy of BiLSTM4DPS, particularly under scenarios with limited account activity data. The results demonstrate that BiLSTM4DPS achieves remarkable predictive accuracy, surpassing existing state-of-the-art methods.