Graph neural networks (GNNs) for fraud detection has received extensive attention, where malicious behaviors often exhibit complex relational patterns. Despite their success, the GNN architecture design of existing graph-based fraud detection methods requires significant manual work and expert knowledge. The application of manually designed architectures to diverse real-world scenarios remains a huge time cost, as it requires numerous parameter tuning for varying conditions. Moreover, the GNN-based methods suffer from the over-smoothing problem during multi-layer message passing, which limits the performance in the fraud detection task. To address these problems, we propose the Automatic lightning decoupled Graph neural architecture search for Fraud Detection (AutoGFD). Specifically, AutoGFD designs the decoupled search algorithm to automatically construct the optimal architecture from the specialized architecture search space for the fraud detection task, which can effectively solve the over-smoothing problem. In addition, AutoGFD designs the lightning search tuning mechanism to improve the efficiency of architecture estimation. As far as we know, AutoGFD is the first attempt to design decoupled architecture search for fraud detection, which can automatically search for optimal architectures in different fraud detection scenarios without manual design and expert knowledge. The experimental results based on multiple benchmark datasets show that AutoGFD can achieve significant performance advantages over state-of-the-art baseline methods.

Graph fraud detection has garnered significant attention as Graph Neural Networks (GNNs) have proven effective in modeling complex relationships within multimodal data. However, existing graph fraud detection methods typically use preprocessed node embeddings and predefined graph structures to reveal fraudsters, which ignore the rich semantic cues contained in raw textual information. Although Large Language Models (LLMs) exhibit powerful capabilities in processing textual information, it remains a significant challenge to perform multimodal fusion of processed textual embeddings with graph structures. In this paper, we propose a Multi-level LLM Enhanced Graph Fraud Detection framework called MLED. In MLED, we utilize LLMs to extract external knowledge from textual information to enhance graph fraud detection methods. To integrate LLMs with graph structure information and enhance the ability to distinguish fraudsters, we design a multi-level LLM enhanced framework including type-level enhancer and relation-level enhancer. One is to enhance the difference between the fraudsters and the benign entities, the other is to enhance the importance of the fraudsters in different relations. The experiments on four real-world datasets show that MLED achieves state-of-the-art performance in graph fraud detection as a generalized framework that can be applied to existing methods.

Heterogeneous temporal graphs (HTGs) are ubiquitous data structures in the real world. Recently, to enhance representation learning on HTGs, numerous attention-based neural networks have been proposed. Despite these successes, existing methods rely on a decoupled temporal and spatial learning paradigm, which weakens interactions of spatio-temporal information and leads to a high model complexity. To bridge this gap, we propose a novel learning paradigm for HTGs called Simple and Efficient Heterogeneous Temporal Graph Neural Network (SE-HTGNN). Specifically, we innovatively integrate temporal modeling into spatial learning via a novel dynamic attention mechanism, which substantially reduces model complexity while enhancing discriminative representation learning on HTGs. Additionally, to comprehensively and adaptively understand HTGs, we leverage large language models to prompt SE-HTGNN, enabling the model to capture the implicit properties of node types as prior knowledge. Extensive experiments demonstrate that SE-HTGNN achieves up to 10× speed-up over the state-of-the-art and latest baseline while maintaining the best forecasting accuracy.

In recent years, heterogeneous temporal graphs (HTGs) have attracted substantial attention in applications of information retrieval such as recommender systems and social networks. To enhance representation learning in HTGs, numerous tailored neural networks have recently been proposed. Despite these successes, existing methods adopt independent parameterization strategies to handle various data distributions in HTGs, leading to optimization challenges and speed bottlenecks. To bridge this gap, this paper proposes a novel transformer-based representation learning paradigm for HTGs called HTGformer. Specifically, assisted by two major modules, i.e., a graph embedding layer and a heterogeneous-temporal encoder, HTGformer can effectively and efficiently capture spatio-temporal heterogeneous information in HTGs for comprehensive node representations. Extensive experiments demonstrate that HTGformer achieves up to 6× speed-up compared to the state-of-the-art baseline while maintaining the best forecasting accuracy.

Heterogeneous dynamic graph neural networks (HDGNNs) are effective methods for processing heterogeneous temporal graphs (HTGs), which serve as ubiquitous data structures in real-world scenarios. The previous HDGNN paradigm obtains representations of future target nodes by mining the spatial heterogeneity and temporal dependence of node attributes, ignoring the learning of relation temporal dependence. However, through experience, we find that the learning of relation temporal dependence, which describes the evolving trends in the importance of neighbors under a certain relation, is beneficial for representation learning of HTGs. To bridge this gap, we propose a novel end-to-end heterogeneous temporal graph learning paradigm called Relation Time-aware Heterogeneous Dynamic Graph Neural Networks (ReTag). Compared to previous HDGNNs, ReTag extracts the temporal dependence of relations from historical relation information and the evolving node attributes to drive subsequent spatio-temporal representation learning. As far as we know, ReTag is the first attempt to perform learning of the temporal dependence of relation, which can generate a more effective representation for different downstream tasks of HTGs. The experimental results of different downstream tasks of HTGs based on multiple benchmark datasets show that ReTag can obtain obvious performance advantages compared with the sota baseline method.

To alleviate the over-smoothing problem caused by deep graph neural networks, decoupled graph neural networks (DGNNs) are proposed. DGNNs decouple the graph neural network into two atomic operations, the propagation (P) operation and the transformation (T) operation. Since manually designing the architecture of DGNNs is a time-consuming and expert-dependent process, the DF-GNAS method is designed, which can automatically construct the architecture of DGNNs with fixed propagation operation and deep layers. The propagation operation is a key process for DGNNs to aggregate graph structure information. However, DF-GNAS automatically designs DGNN architecture using fixed propagation operation for different graph structures will cause performance loss. Meanwhile, DF-GNAS designs deep DGNNs for graphs with simple distributions, which may lead to overfitting problems. To solve the above challenges, we propose the Decoupled Graph Neural Architecture Search with Variable Propagation Operation and Appropriate Depth (DGNAS-PD) method. In DGNAS-PD, we design a DGNN operation space with variable efficient propagation operations in order to better aggregate information on different graph structures. We build an effective genetic search strategy to adaptively design appropriate DGNN depths instead of deep DGNNs for the graph with simple distributions in DGNAS-PD. The experiments on five real-world graphs show that DGNAS-PD outperforms state-of-art baseline methods.

In recent years, heterogeneous graph neural networks (HGNNs) have achieved excellent performance. The efficient HGNNs consist of meta-graphs and aggregation operations. Since manually designing meta-graph is an expert-dependent and time-consuming process, the performance of HGNNs is limited. To address this challenge, the differentiable meta-graph search has been proposed to obtain promising meta-graph automatically. However, the previous differentiable meta-graph search constructs the supernet without learnable aggregation operations, which limits the semantics extracting ability of HGNNs with automatically designed meta-graph for downstream tasks. To solve this problem, we propose the Meta-graph Search with Learnable Supernet for Heterogeneous Graph Neural Networks (MSLS). Specifically, to obtain better performance HGNNs, MSLS constructs a supernet with learnable aggregation operations based on the meta-graphs. MSLS adopts decoupling training to train the learnable supernet and obtains the optimal meta-graph with learnable aggregation operations using a constrained evolution strategy. Extensive experiments show that our method (MSLS) achieves the best performance in different tasks.