Artificial intelligence (AI) is reshaping how digital content is created, circulated, and interpreted, making conventional credibility cues less reliable in computational social systems. This article examines associations among AI-generated content (AIGC), perceived trust, detection confidence, and cybersecurity vulnerability. Specifically, we proposed a comprehensive trust assessment framework to quantify detection confidence metrics, content authentication, user behavioral patterns, and security threat vectors. The experimental results demonstrate trust degradation across content types, detection performance for linguistic, metadata, and behavioral signals, and elevated attack success rates for synthetic-content conditions. The results show that AIGC is associated with lower user trust and higher security vulnerability, while also revealing important limits in causal interpretation, generalization across modalities, and deployment-scale defense evaluation.
Multimodal recommendation systems exploit heterogeneous item modalities to improve recommendation performance. However, existing graph-based and contrastive learning approaches still suffer from noisy semantic graph construction, error propagation caused by graph augmentation, and substantial computational overhead. To overcome these challenges, we propose a novel multimodal recommendation framework, namely multidomain latent semantic mining with augmentation-free contrastive learning (MSACL), which enhances the quality of item-item semantic graphs while learning robust user and item representations. Specifically, MSACL first constructs multidomain item-item semantic graphs (MISGs) from the spatial, wavelet, and interaction domains to capture complementary semantic relationships. These graphs are subsequently fused to generate an item topological graph with reliable structural information. Furthermore, we design an augmentation-free graph contrastive learning (AGCL) strategy that leverages intrinsic contrastive views naturally formed across adjacent graph neural network (GNN) layers. Each node and its corresponding neighbors in the previous layer are regarded as positive pairs, whereas all other nodes are treated as negative samples. This design effectively avoids the noise introduced by conventional graph augmentation and eliminates the costly process of generating additional graph views. Consequently, MSACL reduces computational overhead while enabling more efficient and robust representation learning. Extensive experiments on three public datasets demonstrate that MSACL consistently outperforms state-of-the-art multimodal recommendation methods, validating the effectiveness of integrating multidomain semantic graph modeling with augmentation-free contrastive learning for user preference learning and item representation enhancement. Our code is available at https://github.com/cqjtumathai/MSACL.
Aspect-based sentiment analysis (ABSA) aims to determine the emotional orientation toward particular aspect terms, constituting a fine-grained task in natural language processing (NLP). Most prior approaches rely on homogeneous graph neural networks, which underutilise the heterogeneity of node types and frequently depend on a static syntactic representation, limiting the capacity of the model to incorporate syntactic structures. To mitigate these, this research proposes an adaptive syntax with actor–critic in a heterogeneous graph attention network (ADS-HAN). The model constructs a heterogeneous graph containing three node types and introduces a relation-type attention utilized to integrate heterogeneous information. To improve the adaptability of syntactic representations, we incorporate an advantage actor–critic algorithm to adaptively select the optimal syntactic matrix. To achieve a synergistic fusion of syntax with semantics, the model uses cross-attention for fusing syntactic with semantic features, which are ultimately used for sentiment classification. Experiments conducted in public datasets demonstrate that our approach achieves current main methods on both accuracy and macro-F1 metrics; ablation studies and sensitivity tests provide further validation of the effectiveness of each module. By integrating ADS with a HAN, the effectiveness and robustness of ABSA can be substantially enhanced.
Intelligent transportation is a significant concern of the computational social system; multitarget multicamera vehicle tracking (MTMCT) has become crucial for intelligent traffic supervision and urban safety management. However, maintaining spatiotemporal consistency of vehicle objects poses a significant challenge in complex urban road scenarios. To address this issue, we propose a multiobject vehicle tracking framework based on spatiotemporal constraints and regional semantics by integrating several complementary components. First, the spatiotemporal consistency restoration (STCR) module leverages the assumption of short-term spatial and appearance consistency of vehicles, integrating intersection over union (IoU) and re-identification (ReID) feature metrics to enhance target association and mitigate ID switches caused by occlusion and prolonged static scenes. Second, the framework presents a multicamera tracking optimization method based on a lane-level partitioning (LPMT) module to reduce the computational complexity of cross-camera target matching by introducing a fine-grained lane constraint mechanism as well as providing effective spatial constraints to filter unreasonable matching results. Next, a region semantic adaptive matching (RSAM) strategy refines each lane into entry, middle, and exit zones to encode camera-handover regularities, and couples this region prior with the lane constraint in a hierarchical matching framework that rejects semantically implausible cross-camera transitions. Finally, extensive experiments on the AI CITY Challenge S02 dataset demonstrate that this method improves the IDF1 and multiple object tracking accuracy metrics by 6.49% and 3.51%, respectively, compared with the baseline. These results verify the effectiveness of the proposed method in addressing MTMCT challenges in complex scenarios.
V2X-enabled intelligent transportation systems are representative computational social systems in which vehicles, roadside units, cloud platforms, and human operators interact through edge-managed communication infrastructures. In such environments, cryptographic anomalies in encrypted traffic affect not only communication security but also operator perception of service trustworthiness and intervention priority. However, cryptographic negotiation metadata is often unavailable because of partial captures, tunneling, encapsulation, or proprietary protocol stacks, which makes ciphertext-only anomaly assessment necessary. Existing deep models offer limited transparency for operator-facing auditing, whereas conventional fuzzy neural networks scale poorly on high-dimensional ciphertext fingerprints. To address this challenge, this article presents ensemble of evolutionary quantum fuzzy neural network (EEQFNN), an explainable quantum fuzzy intelligence framework for perception-aware cryptographic anomaly assessment in V2X-enabled computational social systems. EEQFNN combines global quanvolution for compact feature enhancement with a circuit-centric quantum fuzzy neural network (QFNN) equipped with a quantum T-norm operator for efficient high-dimensional rule activation on a quantum simulator. An offline evolutionary design based on the non-dominated sorting genetic algorithm II (NSGA-II) balances primitive-identification accuracy and deployment complexity, while an elite ensemble improves robustness under heterogeneous traffic conditions. A policy-aware postinference layer further converts primitive predictions into compliance scores, anomaly severity, confidence estimates, and rule-level evidence for operator-facing assessment. Experiments on a scenario-driven, policy-aware V2X encrypted communication testbed show that EEQFNN improves primitive-identification accuracy by 2.35 percentage points over the best-performing baseline while enabling interpretable and policy-aware anomaly assessment at the edge.
