Large Language Model-based Multi-Agent Systems (LLM-MAS) have attracted significant attention due to their advantages in handling complex tasks. Current research primarily focuses on task-specific agent design, cooperation mechanisms, and pipeline optimization. However, the robustness of LLM-MAS remains underexplored. Internal conflicts and insufficient information sharing among agents can expose the system to global-level failures, especially under abnormal conditions or adversarial attacks. To address these challenges, we propose a generalizable and computationally efficient protection mechanism, RollMAS. First, we design the action negotiation algorithm that mitigates risks arising from agent discrepancies by enabling multiple local policies to converge through signal synchronization in the vector space. Second, the sharing redundancy enhancement algorithm optimizes MAS robustness by maximizing an approximate natural eigenvalue of the corresponding adjacency matrix, facilitating resilient and efficient information sharing with controlled communication overhead. Extensive experiments on traffic control and question answering tasks, spanning 7 datasets and 13 baselines, demonstrate that our methods substantially enhance the robustness and effectiveness of LLM-MAS.

Graph anomaly detection attracts considerable interest across a variety of application domains, including fraud detection within social networks, identifying money laundering activities in transaction graphs, etc. The advent of Graph Neural Networks (GNNs) has enhanced existing deep learning methods to capture the anomaly patterns in the latent space and achieve satisfactory results. However, the performance deficiency of current GNNs is primarily caused by the scarce labeled anomalies in a specific real-world application. Unlike the total neglect of valuable labeled anomalies in unsupervised approaches or the potential overfitting in supervised approaches, we proposed a few-shot-oriented framework GUDI in this paper. GUDI is the first work to incorporate a self-supervised pre-training approach to capture general graph patterns across domains and design a classifier with few-shot learning to model the labeled data within a specific domain. The proposed guided diffusion mechanism synthesizes the outcomes of the pre-trained model and the domain-specific classifier in the inference phase, which negates the need for fine-tuning the large pre-trained model, thereby facilitating efficient domain adaptation. GUDI retains the ability to uncover unknown anomalies through unsupervised pre-training while also possessing the ability to identify known anomaly natterns from few-shot labels.

In today's fast-paced world, advertisers are increasingly demanding real-time and accurate personalized ad delivery based on dynamic preference modeling, which emphasizes the temporality existing in both user preference and product characteristics. Meanwhile, with the development of graph neural networks (GNNs), E-commerce knowledge graphs (KG) with rich semantic relatedness are invoked to improve accuracy and provide appropriate explanations to encourage advertisers' willingness to invest in ad expenses. However, it is still challenging for existing methods to comprehensively consider both time-series interactions and graph-structured knowledge triples in a unified model, i.e., the case in knowledge-aware dynamic advertising. The interaction graph between users and products changes rapidly over time, while the knowledge in KG remains relatively stable. This results in an uneven distribution of temporal and semantic information, causing existing GNNs to fail in this scenario. In this work, we quantitatively define the above phenomenon as temporal unevenness and introduce the Incremental Leveling Network (InLN) with three novel techniques: the periodic-focusing window for node-level dynamic modeling, the biased temporal walk for subgraph-level dynamic modeling and the incremental leveling mechanism for KG updating. Verified by comprehensive and intensive experiments, InLN outperforms nine baseline models in three tasks by substantial margins, reaching up to a 9.9% improvement and averaging a 5.7% increase.

With the upsurge of online banking, mobile payment, and virtual currency, new money-laundering crimes easily conceal in the enormous transaction volume. The traditional rule-based methods with large amounts of alerting thresholds are already incapable of handling the fast-changing transaction networks. Recently, the DL models represented by the graph neural networks (GNNs) show the potential to capture money-laundering modes with high accuracy. However, most related works are still far from practical deployment in the industry. Based on our practice at WeBank, there are three major challenges: Firstly, supervised learning is infeasible facing the extraordinarily large-scale but imbalanced data, with hundreds of millions of active accounts but only thousands of anomalies. Secondly, the real-world transactions form a sparse network with millions of isolated user groups, which overflows the expressive ability of current node-level GNNs. Thirdly, the explanation for each suspicious account is mandatory by the government for double check, which conflicts with the black-box nature of most DL models. Therefore, we proposed Diga, the first work to apply the diffusion probabilistic model to a graph anomaly detection problem with three novel techniques: the biased K-hop PageRank, the semi-supervised guided diffusion and the novel weight-sharing GNN layer. The effectiveness and efficiency of Diga are verified via intensive experiments on both industrial and public datasets.

Recently, knowledge-graph-enhanced recommendation systems have attracted much attention, since knowledge graph (KG) can help improving the dataset quality and offering rich semantics for explainable recommendation. However, current KG-enhanced solutions focus on analyzing user behaviors on the product level and lack effective approaches to extract user preference towards product category, which is essential for better recommendation because users shopping online normally have strong preference towards distinctive product categories, not merely on products, according to various user studies. Moreover, the existing pure embedding-based recommendation methods can only utilize KGs with a limited size, which is not adaptable to many real-world applications. In this paper, we generalize the recommendation problem with preference mining as a compound knowledge reasoning task and propose a novel multi-agent system, called Mcore, which can promote model performance by mining users’ high-level interests and is adaptable to large KGs. Specifically, we split the overall problem and allocate sub-task to each agent: Coordinate Agent takes charge of recognizing the product-category preference of current user, while Relation Agent and Entity Agent perform KG reasoning cooperatively from a user node towards the preferred categories and terminate at a product node as recommendation. To train this heterogeneous multi-agent system, where agents own various functionalities, we propose an asynchronous reinforcement training pipeline, called Multi-agent Collaborative Learning. The extensive experiments on real datasets demonstrate the effectiveness and adaptability of Mcore on recommendation tasks.