Money laundering remains a critical challenge for financial systems because of the complex, hidden, and interlinked nature of illicit financial transaction networks. Understanding how these networks respond to targeted disruption is essential for exposing structural vulnerabilities and refining existing anti-money laundering (AML) prevention and intervention strategies. This study involves a social network analysis (SNA)-based resilience framework to evaluate the robustness of financial transaction networks through targeted node removal. In this approach, a network is represented as a directed graph, where nodes correspond to accounts and edges represent transactions. Centrality measures (i.e., degree, closeness, betweenness and pagerank), which capture local influence, global reach, and control over information flow, are applied to identify the most influential nodes. Network resilience is assessed by analyzing the variation in the size of the Largest Connected Component (LCC) under progressive node removal. An adaptive LCC-based resilience strategy is used, starting with large batches of nodes and gradually moving to smaller ones until the LCC drops below 50% of its original size, allowing for a more detailed analysis near the fragmentation threshold. The findings reveal that Betweenness centrality is the most effective metric in disrupting network connectivity under targeted attack scenarios, both outflow- and inflow-based analyses. Specifically, targeting only the top 2% of nodes by Betweenness centrality collapses the network’s core, reducing the Largest Connected Component (LCC) to 60% of its original size. In contrast, random attack strategy exhibit limited impact on overall network resilience compared to targeted approaches. Our findings provide actionable AML insights, showing that resilience-driven targeting of structurally critical accounts can effectively fragment money laundering networks and support more focused interdiction strategies.

Graph neural networks (GNNs) combined with contrastive learning have shown promise in mitigating data sparsity and noise in recommender systems. However, current approaches face two key challenges: (1) they often conflate genuine user preferences with group-level influences, introducing confounding bias in representation learning; (2) conventional graph contrastive methods rely on random perturbations that can compromise semantic structure and amplify noise. To overcome these issues, we propose CaDCR - Causal-Augmented Disentanglement for Contrastive Recommendation. CaDCR introduces a dual-channel mechanism to separate individual and collective intent, applying causal interventions to eliminate bias from factors like popularity. Additionally, it employs dual contrastive regularization in both interaction and intent spaces, preserving semantic integrity and enhancing representation quality without manual augmentation. Experiments on Gowalla and other datasets show CaDCR outperforms 14 mainstream methods on metrics like Recall and NDCG, validating its effectiveness and practical relevance.

The rapid spread of fake news on social media seriously threatens the orientation of public opinion and public trust. Although existing fake news detection methods have made significant progress in specific scenarios, there are still open challenges such as strong dependence on labeled data and difficulty in capturing the structure of news propagation. To address these challenges we propose GUDSAE, a graph up-down sampling adaptive en semble method for fake news detection. Firstly, we model the interaction process between news and users as a propagation graph, and we adopt random node feature masking to enhance input diversity. Secondly, we intro duce a graph autoencoder (GAE) architecture composed of GCN, graph pooling (gPool) layer and graph unpooling (gUnpool) layer for graph up-down sampling, which is effective in both extracting multi-layer topologies and se mantic embeddings under unsupervised conditions, as well as generating pseudo-labels on the basis of structural similarity for supervised optimization. Finally, we design an adaptive ensemble strategy of local feature fusion and global information optimization. Here, the features and assigned labels were dynamically fused respectively based on the reconstruction error and classification consistency to construct an end-to-end fake news detection framework. Experiments on the Politifact and GossipCop datasets show that GUDSAE performs significantly better than its competitors with an accuracy improvements of 1.92% and 2.84% respectively.

Accurately identifying key nodes in complex networks is crucial for understanding propagation dynamics and optimizing resource allocation. However, existing methods often suffer from long-distance noise interference and rigid distance measures, which reduce identification accuracy. This paper proposes a Dual-Constrained and Adaptive Gravity (DCAG) method that integrates structural constraints and adaptive distance into the gravity framework for key node identification. First, a dual structural constraint combining radius truncation and non-decreasing kernel values refines gravitational interactions within the second-order neighborhood, filtering long-distance noise and reducing computational cost. Second, an adaptive effective structural distance with degree penalty terms captures the shortcut effect of hub nodes, so that influence attenuation better matches actual propagation patterns. Third, a normalized gravitational factor based on the average kernel value eliminates manual parameter tuning and balances gravity distribution across different core levels. Extensive experiments on nine real networks show that DCAG achieves the best single-node discrimination and the highest correlation with the SIR propagation model among competing methods. On a million-node network, DCAG reduces runtime to 0.48 seconds, demonstrating sub-second response and quasi-linear scalability. In addition, when used as a general enhancement framework, DCAG significantly improves the performance of traditional centrality indicators, highlighting its universality and robustness.

Graph Contrastive Learning (GCL) has recently emerged as a powerful paradigm for recommendation systems. However,its practical adoption is hindered by two critical challenges: (a) The representational capacity and interpretability of recommendation models; (b) The inherent flaws of data augmentation strategies in intent decoupling learning, which often introduce misleading self-supervised signals due to noise. To address these issues, we propose a novel framework that integrates adaptive augmentation with intent-aware modeling to improve the challenge (a), called AIARec. Our approach addresses the sparsity of implicit feedback in bipartite graphs by introducing a Gaussian distribution-based graph generation strategy for robust node feature encoding. Furthermore, we design an adaptive feature-level noise perturbation mechanism, governed by an embedding table, which judiciously guides the reconstruction of the bipartite graph using latent intent information. This mechanism not only mitigates excessive noise perturbation but also accentuates the intrinsic intention features of users and items, thereby strengthening the challenge (b). To further refine the learned representations, we develop a two-domain aware graph contrastive learning framework that optimizes the consistency and uniformity of node embeddings across multiple domains. Extensive experiments on real datasets (e.g., Yelp, AmazonBook) show that AIARec outperforms 16 state-of-the-art baselines (e.g., BIGCF, LightGCN) on Recall and NDCG in four metrics (recall, NDCG, etc.). By explicitly modeling user-item interactions through interpretable intent factors, AIARec advances both the performance and explainability of GCL-based recommender systems, offering a principled solution to noisy augmentation and sparse interaction challenges.

Contemporary techniques for identifying key nodes in complex networks typically rely on the static topology of the network, often neglecting the potential dynamic information available. We introduce a novel centrality measurement approach named gravity box-coverage and effective distance (GBED). It capitalizes on the notion that the internal structure of the gravity box encapsulates crucial information about nodes. It transforms static Euclidean distance into dynamic effective distance (ED), extracting concealed insights through an analysis of both static and dynamic topological paths. Initially, the ED between nodes is computed based on node arrival probabilities. Subsequently, the box-coverage algorithm defines the influence area of nodes. The improved gravity model is then applied to estimate the interaction ability between nodes. Finally, the local influence capability score of the node’s box, covering the influence region, is calculated. The global influence capability score of the node is aggregated according to the neighborhood rule. We compare it with five established methods based on nine real-world networks. In the SIR epidemic spreading, the nodes identified by GBED exhibit a broader range of influence, and the correlation between estimated influences of nodes from GBED and real influences by simulation is higher than correlations associated with other algorithms.

Predicting links between groups accurately, especially when dealing with brief or emerging connections, remains a significant challenge in network analysis. We propose a Contrast View Attention Network (CVAN) to address this issue. CVAN leverages attention mechanisms and contrastive learning techniques to enhance the accuracy of group link predictions. We employ a graph neural network (GNN)-based encoder to capture the heterogeneity of nodes and edges, facilitating the modeling of interactions between group members. CVAN learns multi-view group representations, capturing diverse member weights and group characteristics from various perspectives. In addition, CVAN integrates infoNCE loss and Triplet loss functions to optimize the prediction process and establish a comprehensive group comparison view. We enhance the embedded representations of groups and individuals by combining the GNN encoder with a multilayer perceptron (MLP). Moreover, employing a cross-view contrast learning method, CVAN constructs a group contrast loss optimization function, improving the efficiency of group-to-group link prediction. We evaluate the performance of CVAN on three datasets from different domains. Our experiments compare CVAN with 11 popular methods. On the Weeplaces-G dataset, CVAN outperformed the latest MMAN method by 1.5% on MRR and HITS@10, and by 3.3% on NDCG@10. On the MAG and Aminer-G datasets, CVAN achieved improvements ranging from 1.5% to 3% on MRR and HITS@10, along with a 3% gain on NDCG@10, further demonstrating its accuracy and robustness. These results showcase CVAN’s superior capability in accurately predicting links between groups and underscore its effectiveness across various group link prediction tasks, highlighting its strong potential for real-world applications.

With the development of deep learning and other technologies, the research of information propagation prediction has also achieved important research achievements. However, the existing information diffusion studies either focus on the attention relationships of users or they predict the information according to the diffusion relationships of users, which makes the prediction results have certain limitations. Therefore, a prediction model has been proposed spatial–temporal attention heterogeneous graph convolutional networks (STAHGCNs). First, we use GCN to learn user influence relationships and user behavior relationships, and we propose a user representation fusion mechanism to learn the user characteristics. Second, to account for the dynamics of user behavior, a temporal attention mechanism strategy is used to encode time into the heterogeneous graph to obtain a more expressive user representation. Finally, the obtained user representation is input into the multihead attention mechanism for information propagation prediction. Experimental results performed on the Twitter, Douban, Digg, and Memetracker datasets have shown that the proposed STAHGCN model increased by 8.80% and 6.74% at hits@N and map@N, respectively, which are significantly better than the original latest DyHGCN model. The proposed STAHGCN model effectively integrates spatial factors, such as time factor, user influence, and behavior, which greatly improves the accuracy of information propagation prediction and has great significance for rumor monitoring and malicious account detection.

Despite the huge importance that the centrality metrics have in understanding the topology of a network, too little is known about the effects that small alterations in the topology of the input graph induce in the norm of the vector that stores the node centralities. If so, then it could be possible to avoid re-calculating the vector of centrality metrics if some minimal changes occur in the network topology, which would allow for significant computational savings. Hence, after formalising the notion of centrality, three of the most basic metrics were herein considered (i.e., Degree, Eigenvector, and Katz centrality). To perform the simulations, two probabilistic failure models were used to describe alterations in network topology: Uniform (i.e., all nodes can be independently deleted from the network with a fixed probability) and Best Connected (i.e., the probability a node is removed depends on its degree). Our analysis suggests that, in the case of degree, small variations in the topology of the input graph determine small variations in Degree centrality, independently of the topological features of the input graph; conversely, both Eigenvector and Katz centralities can be extremely sensitive to changes in the topology of the input graph. In other words, if the input graph has some specific features, even small changes in the topology of the input graph can have catastrophic effects on the Eigenvector or Katz centrality.

The problem of detecting communities in real-world networks has been extensively studied in the past, but most of the existing approaches work on single-domain networks, i.e. they consider only one type of relationship between nodes. Single-domain networks may contain noisy edges and they may lack some important information. Thus, some authors have proposed to consider the multiple relationships that connect the nodes of a network, thus obtaining multi-domain networks. However, most community detection approaches are limited to multi-layer networks, i.e. networks generated from the superposition of several single-domain networks (called layers) that are regarded as independent of each other. In addition to being computationally expensive, multi-layer approaches might yield inaccurate results because they ignore potential dependencies between layers. This paper proposes a multi-domain discrete-time quantum walks (MDQW) model for multi-domain networks. First, the walking space of network nodes in multi-domain network is constructed. Second, the quantum permutation circuit of the coin state is designed based on the coded particle state. Then, using different coin states, the shift operator performs several quantum walks on the particles. Finally, the corresponding update rule is selected to move the node according to the measurement result of the quantum state. With continuous update iteration, the shift operator automatically optimizes the discovered community structure. We experimentally compared our MDQW method with four state-of-the-art competitors on five real datasets. We used the normalized mutual information (NMI) to compare clustering quality, and we report an increase in NMI of up to 3.51 of our MDQW method in comparison with the second-best performing competitor. The MDQW method is much faster than its competitors, allowing us to conclude that MDQW is a useful tool in the analysis of large real-life multi-domain networks. Finally, we illustrate the usefulness of our approach on two real-world case studies.

Traditional community detection models either ignore the feature space information and require a large amount of domain knowledge to define the meta-paths manually, or fail to distinguish the importance of different meta-paths. To overcome these limitations, we propose a novel heterogeneous graph community detection method (called KGNN_HCD, heterogeneous graph Community Detection method based on K-nearest neighbor Graph Neural Network). Firstly, the similarity matrix is generated to construct the topological structure of K-nearest neighbor graph; secondly, the meta-path information matrix is generated using a meta-path transformation layer (Mp-Trans Layer) by adding weighted convolution; finally, a graph convolutional network (GCN) is used to learn high-quality node representation, and the k-means algorithm is adopted on node embeddings to detect the community structure. We perform extensive experiments and on three heterogeneous datasets, ACM, DBLP and IMDB, and we consider as competitors 11 community detection methods such as CP-GNN and GTN. The experimental results show that the proposed KGNN_HCD method improves 2.54% and 2.56% on the ACM dataset, 2.59% and 1.47% on the DBLP dataset, and 1.22% and 1.67% on the IMDB dataset for both NMI and ARI. Experiments findings suggest that the proposed KGNN_HCD method is reasonable and effective, and KGNN_HCD can be applied to complex network classification and clustering tasks.

The improvement of positioning accuracy in Wireless Sensor Networks (hereafter, WSN) is crucial to develop advanced Internet of Things (IOT, for short) applications. However, the conventional distance vector-hop (DV-Hop) localization algorithm has shortcomings such as low accuracy and weak stability. To overcome these shortcomings, this paper proposes a hybrid improved compressed particle swarm optimization algorithm (HICPSO), which consists of a scheme of linearly decreasing inertia weights, compressed velocity vectors, population Gaussian variants and optimal boundary selection. Then, HICPSO is integrated with DV-Hop to gradually reduce the distance error of least squares method (LSM) estimated with the efficient search advantage of HICPSO. Our simulation results show that the HICPSO algorithm possesses better computational accuracy and search performance on the 22 benchmark test functions compared with the algorithms such as the Improved Adaptive Genetic Algorithm (IAGA) and Adaptive Weighted Particle Swarm Optimizer (AWPSO). Meanwhile, compared with IAGA and AWPSO, the positioning accuracy of HICPSO-based positioning algorithm is improved by 4.28% and 4.76% respectively, and the stability is improved by one order of magnitude.

Standard machine learning and deep learning architectures have been widely used in the field of sentiment analysis, but their performance is unsatisfactory if the input texts are short (e.g., social media posts). Specifically, the accuracy of standard machine learning methods crucially depends on the richness and completeness of the features used to represent the texts, and in the case of short messages, it is often difficult to obtain high-quality features. Conversely, methods based on deep learning can achieve better expressiveness, but these methods are computationally demanding and often suffer from over-fitting. This paper proposes a new adaptive evolutionary computational integrated learning model (AdaECELM) to overcome the problems encountered by traditional machine learning and deep learning models in sentiment analysis for short texts. AdaECELM consists of three phases: feature selection, sub classifier training, and global integration learning. First, a grid search is used for feature extraction and selection of term frequency-inverse document frequency (TF-IDF). Second, cuckoo search (CS) is introduced to optimize the combined hyperparameters in the sub-classifier support vector machine (SVM). Finally, the training set is divided into different feature subsets for sub-classifier training, and then the trained sub-classifiers are integrated and learned using the AdaBoost integrated soft voting method. Extensive experiments were conducted on six real polar sentiment analysis data sets. The results show that the AdaECELM model outperforms the traditional ML comparison methods according to evaluation metrics such as accuracy, precision, recall, and F1-score in all cases, and we report an improvement in accuracy exceeding 4.5%, the second-best competitor.

Most existing cross-domain recommendation (CDR) systems apply the embedding and mapping idea to tackle the cold-start user problem and, to this end, they learn a common bridge function to transfer the user preferences from the source domain into the target domain. However, sharing a bridge function for all users inevitably leads to biased recommendations. This paper proposes a novel method, named CDR to cold-start users via categorized preference transfer (CDRCPT), to overcome the shortcomings of existing approaches. First, the embeddings of users and items in both the source and target domain are learned through pretraining and we utilize preference encoder to obtain the preference embeddings of users in the source domain. Second, mini-batch clustering is applied in the source domain to group users according to their preferences; here, each cluster identifies a specific class of users, and each cluster is represented by its center. Finally, the general representation is fed into a meta network to learn a bridge function for each available class of users. Experiments on two real data sets show that our CDRCPT method is effective in improving the accuracy and robustness of recommendations.

Traditional methods for influential node identification usually require time consuming network traversal to select the candidate node set. In this article we propose a new influence nodes identification method, called Community-based Backward Generating Network (CBGN). First, the influence maximization framework is built by integrating community detection and Backward Generation Network (BGN); then, nodes in each community are selected using a new method, called imp_BGN, that uses graph traversal to assist the construction of BGN. The ultimate goal of the network generation method is to find a sequence of nodes that can minimize the cost function, and to select high influential nodes without restoring the original network during network construction. finally, an improved submodular CELF (Cost Effective Lazy Forward) algorithm is proposed to hunt for the final seed node from the candidate node pool considering the location relation and structural similarity among nodes. Experimental results show that: in the SIR (susceptible-infected-recovered) model experiment, compared with the benchmark methods, the infection scale of the proposed CBGN method in 6 real networks is improved by 0.45%, 0.59%, 0.84%, 1.05%, 0.71% and 0.14%, respectively.

In view that the K-shell decomposition method can only effectively identify a single most influential node, but cannot accurately identify a group of most influential nodes, this article proposes a hybrid method based on K-shell decomposition to identify the most influential spreaders in complex networks. First, the K-shell decomposition method is used to decompose the network, and the network is regarded as a hierarchical structure from the inner core to the periphery core. Second, the existing centrality methods such as H-index are used as the secondary score of the proposed method to select nodes in each hierarchy of the network. In addition, for the sake of alleviating the overlapping problem, the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) method is introduced to calculate the comprehensive score of secondary score and overlapping range, and the node with the highest comprehensive score will be selected in each round. The proposed algorithm can be used as a general framework to improve the existing centrality method which can represent nodes with definite values of centrality. Experimental results show that in the susceptible–infected–recovered (SIR) model experiment, compared with the benchmark methods, the infection scale of the proposed K-TOPSIS method in nine real networks is improved by 1.15%, 2.23%, 1.95%, 3.12%, 6.29%, −0.37%, 4.01%, 0.48%, and 0.48%, respectively. The novel method is improved by 0.44, 1.18, 1.16, 11.30, 2.03, 2.53, 2.70, and 2.13 in average shortest path length experiment, respectively, except for Facebook network. It shows that the novel method is reasonable and effective.

Social Network Analysis (SNA) is an interdisciplinary science that focuses on discovering the patterns of individuals interactions. In particular, practitioners have used SNA to describe and analyze criminal networks to highlight subgroups, key actors, strengths and weaknesses in order to generate disruption interventions and crime prevention systems. In this paper, the effectiveness of a total of seven disruption strategies for two real Mafia networks is investigated adopting SNA tools. Three interventions targeting actors with a high level of social capital and three interventions targeting those with a high human capital are put to the test and compared between each other and with random node removal. Similar tests on artificial model networks have also been carried out. Simulations show that actor removal based on social capital proves to be the most effective strategy, by leading to the total disruption of the criminal network in the least number of steps. The removal of a specific figure of a Mafia family such as the Caporegime seems also promising in the network disruption.

Road (or street) networks refer to interconnected systems of streets, roads, and pathways which constitute the transportation infrastructure within a city or urban area. These networks play a crucial role in urban planning and design, as they influence various aspects of city life. There are several key elements to be considered in the study of street networks: hierarchy, grid and patterns, connectivity, walkability, transit integration, traffic flow, land use and zoning, and smart mobility solutions. In this study, we use social network analysis tools to identify the most effective strategies to simulate and assess the impact of road network disruption. Our scenario has been constructed by building a road network based on a portion of the geographic map data of Messina (Italy) and implementing random and targeted attacks based on edge centrality measures and street types. The impact of our strategies on the road network structure is evaluated in terms of connectivity.

Random walks simulate the randomness of objects, and are key instruments in various fields such as computer science, biology and physics. The counter part of classical random walks in quantum mechanics are the quantum walks. Quantum walk algorithms provide an exponential speedup over classical algorithms. Classical and quantum random walks can be applied in social network analysis, and can be used to define specific centrality metrics in terms of node occupation on single-layer and multilayer networks. In this paper, we applied these new centrality measures to three real criminal networks derived from an anti-mafia operation named Montagna and a multilayer network derived from them. Our aim is to (i) identify leaders in our criminal networks, (ii) study the dependence between these centralities and the degree, (iii) compare the results obtained for the real multilayer criminal network with those of a synthetic multilayer network which replicates its structure.

Target-specific sentiment analysis is an emerging topic in the field of text mining but current approaches to deriving the polarity of a sentence suffer from two main drawbacks: on one hand, we lack of a large and well-curated corpus, and on the other hand, current solutions based on deep learning are particularly vulnerable to the attack of adversarial samples. A novel target-specific sentiment classification method is proposed. Firstly, the method of masking target entities is applied to replace synonyms and insert words randomly; secondly, the target-specific sentiment classification model of adversarial learning is constructed with six baseline models; finally, we combine data enhancement and adversarial learning to construct target-specific sentiment classification model. Experimental results show that Macro-F1 values are improved by 0.30-2.91, 0.88-2.42 and 0.13-1.94% compared to the six baseline models by using Laptop14, Restaurant14 and Twitter original datasets, respectively, using Adversarial learning. Using word-masking data enhancement samples and Adversarial learning from Laptop14, Restaurant14 and Twitter shows that Macro-F1 values are improved by 0.9-2.64, 1.59-3.09 and 0.18-1.71% compared to the six baseline (SC), respectively. Our method can effectively improve the quality of samples, it improves the classification performance and the capability of adversarial samples defense.