Source-linked AI summary
A Review on Graph Neural Network Methods in Financial Applications
Jianian Wang, Sheng Zhang, Yanghua Xiao, Rui Song
TL;DR
Financial graphs are often heterogeneous or time-varying, creating challenges for representing features, relations, and temporal information in financial tasks. This paper reviews GNN methods for finance by categorizing graph types and construction, applications, features, models, code, and research challenges. It provides a systematic resource for understanding, implementing, and developing GNN models across multiple financial tasks.
Problem
Financial data contain complex, heterogeneous, and time-varying relations, while existing GNN surveys underemphasize financial applications and financial surveys lack detailed GNN coverage.
Method
The paper systematically reviews financial GNN methods by categorizing graph types, summarizing node features and methodologies, organizing applications, and discussing challenges and future directions.
Results
The survey organizes financial GNN research across graph categories, five application areas, model information, available code, and five challenges.
Takeaways & Limitations
The review serves as a resource for understanding, implementing, and developing GNN models across multiple financial tasks.
Takeaways & Limitations
Only about 24% of reviewed papers release code, and privacy restrictions limit open financial datasets, hindering reproducibility and methodological comparison.
Abstract
from arXiv · showhide
With multiple components and relations, financial data are often presented as graph data, since it could represent both the individual features and the complicated relations. Due to the complexity and volatility of the financial market, the graph constructed on the financial data is often heterogeneous or time-varying, which imposes challenges on modeling technology. Among the graph modeling technologies, graph neural network (GNN) models are able to handle the complex graph structure and achieve great performance and thus could be used to solve financial tasks. In this work, we provide a comprehensive review of GNN models in recent financial context. We first categorize the commonly-used financial graphs and summarize the feature processing step for each node. Then we summarize the GNN methodology for each graph type, application in each area, and propose some potential research areas.
1 Introduction
Financial data form complex, frequently changing relational graphs, making feature processing, graph construction, and GNN modeling challenging. This survey addresses the gap between general GNN surveys and financial-application reviews by organizing financial graphs, applications, methods, and challenges.
- Financial systems combine many components and sophisticated, frequently updated relations that can be represented with transaction, user-item, and stock relation graphs.
- GNNs are applied to financial tasks such as stock prediction by representing stocks as nodes and inter-company relations as edges.The workflow constructs a graph and feature matrix, feeds them to a GNN, and uses an MLP for price prediction.
- Financial graphs create challenges for feature processing, graph construction, and modeling because data may be numerical or textual while relations are heterogeneous or time-varying.
- Existing GNN surveys emphasize modeling methodology, whereas financial-application surveys do not cover GNN models in detail.
- The survey categorizes financial graphs and applications, summarizes features, graphs, models, and code, and identifies challenges with future directions.Applications include stock movement prediction, loan default risk prediction, e-commerce recommendation, fraud detection, and event prediction.
2 Graph categorization
Graph construction determines how structural information is represented and influences the downstream modeling methodology. The section categorizes financial graphs by construction methods and graph types, with Table 1 listing graphs used in financial tasks.
- Graph construction is essential because the constructed graph type can determine the follow-up modeling methodology.
- The section categorizes financial graphs according to their construction methods and graph types.
- Table 1 presents a comprehensive list of graphs for financial tasks.
2.1 Graph-related definition
The section defines graph representations and several structural graph types used to describe financial relations. These definitions distinguish nodes, edges, directions, node and edge types, relation types, and graph evolution over time.
- A graph G = (V, E) consists of a node set V and an edge set E, where each edge joins two nodes.
- An adjacency matrix A records node-to-node connection status, with binary entries indicating whether an edge exists in an unweighted graph.
- Undirected edges lack orientation and yield a symmetric adjacency matrix, whereas directed edges point from one node to another.
- A bipartite graph divides nodes into two non-empty disjoint sets, with every edge connecting nodes from different sets.
- Homogeneous graphs use one node type and one edge type; heterogeneous graphs do not, while multi-relation graphs have different edge types.
- A dynamic graph is a sequence of graphs G_seq = {G_1, ..., G_T}, each with its own node and edge sets.
2.2 Graph categorization by construction methods
Financial graphs can be constructed from naturally observed relations, knowledge bases, or inferred similarities. Each approach supplies different relational information and presents distinct processing or validation considerations.
- Some financial data are naturally represented as graphs because they contain relations among data objects.Examples include user-relationship graphs and account-device networks.
- Knowledge-based construction extracts company relations such as suppliers, providers, and partners from knowledge bases to build stock relation networks.This adds information to the graph but requires processing complicated structures and large amounts of knowledge-base information.
- Similarity-based construction connects observations when a feature-similarity measure exceeds a threshold, such as cosine similarity of historical stock prices.The approach is easy to implement, but threshold selection and overlap between graph and feature information require justification.
2.3 Graph categorization by graph types
The survey categorizes financial graphs into five types according to node, edge, relational, and temporal characteristics, linking each type to representative financial applications and modeling challenges.
- Financial graphs are grouped into homogeneous, directed, bipartite, multi-relation, and dynamic categories, and a graph may belong to multiple categories.
- 2.3.1 Homogeneous graph: Homogeneous graphs contain one node type and one edge type, giving them a relatively simple structure compatible with most GNN methods.
- 2.3.2 Directed graph: Directed graphs encode asymmetric relationships through oriented edges, but methods designed for semi-definite adjacency matrices may be unsuitable.
- 2.3.3 Bipartite graph: Bipartite graphs connect two node groups exclusively across groups, supporting account-device, user-item, and other interaction networks.
- 2.3.4 Multi-relation graph: Multi-relation graphs use multiple edge types to represent distinct relationships, while dynamic graphs represent changing relations and features as graph sequences.
- 2.3.5 Dynamic graph: Dynamic graphs are difficult to model because nodes and edges may appear or disappear, breaking operations that require fixed-dimensional matrices.
3 Feature processing
Financial node features are commonly sequential numerical data or text, so the survey describes recurrent and language-processing methods that convert them into representations usable by GNNs.
- Financial node features commonly take sequential numerical or textual form, requiring specialized processing before GNN application.
- 3.1 Sequential numerical data: RNN-based methods capture temporal dependencies in sequential numerical features, with LSTM and GRU approaches forming two main lines of work.
- 3.1.1 LSTM based approach: LSTM encodes historical feature sequences through memory cells and gates, and its final hidden state can form a sequential embedding for a GNN.
- 3.1.2 GRU based approach: GRU uses a gating mechanism with fewer parameters, and attention can weight hidden states according to the differing importance of past days.
- 3.2 Textual information: Textual financial information is converted into vectors using word embeddings, recurrent sentence models, attention, or language models such as BERT.
4 Graph neural network models
The review organizes GNN methods by graph structure, covering homogeneous, directed, bipartite, multi-relation, and dynamic settings. These methods aggregate node, edge, neighborhood, or relation-specific information to model increasingly complex financial graphs.
- Homogeneous graph: GCN encodes local graph structure and node features by aggregating neighboring information, while its simple structure makes it a common benchmark.The propagation uses an adjacency matrix with added self-connections and a trainable layer weight matrix.
- Homogeneous graph: GAT extends neighborhood aggregation by assigning different learned weights to neighboring nodes, whose attention coefficients can indicate relative importance.Unlike GCN, GAT accounts for unequal influence among neighbors and is also frequently used as a benchmark.
- Directed graph: Directed graphs can be handled with spatial GNNs such as GAT, while GGNNs use recurrent propagation to encode event sequences and predict subsequent events.GGNN updates event representations through reset and update gates while aggregating structural information at each propagation step.
- Bipartite graph: Bipartite GNN frameworks alternately aggregate edge and neighbor information to update user and item representations.The framework supports user-chosen aggregation functions, attention-based aggregation, and neighbor sampling for user-item networks.
- Multi-relation graph: Multi-relation methods commonly perform subgraph aggregation within each relation and then inter-relation aggregation across relation-specific embeddings.Relation-specific embeddings may be transformed and combined with attention because different relations capture different facets of node characteristics.
5 Application
The review describes GNN applications across stock prediction, loan default prediction, recommender systems, and fraud detection. These applications use graph structure to represent dependencies among financial entities, but face challenges from graph construction, temporal dynamics, noisy interactions, and heterogeneous data sources.
- Stock movement prediction: Stock movement prediction represents stocks as graph nodes so GNNs can model dependencies that independent-stock approaches omit.Researchers construct stock graphs from historical-price correlations or knowledge-base relations such as supplier, customer, partner, and shareholder links.
- Stock movement prediction: Stock prediction remains difficult because no off-the-shelf inter-stock graph exists and financial data require temporal and multi-source modeling.Relevant sources include financial statements, news, and pricing information, while stock relations may evolve over time.
- Loan default prediction: Loan default prediction uses graphs of borrower interactions, including directed guarantee networks and user relations based on social connections, transactions, or device usage.These graph structures extend beyond individual borrower features to represent influences among related individuals.
- Recommender system: Recommender-system applications use user-item bipartite graphs to model users, products, and their interactions for preference prediction and item recommendation.GNN methods are also applied to click-rate prediction and fake-review detection, including directed and signed comment graphs.
- Fraud detection: Fraud detection formulates abnormal user connectivity as a graph node-classification problem across payment fraud, identity theft, scams, and insurance fraud.The graph formulation reflects the observation that fraudsters may exhibit abnormal connectivity with other users.
6 Challenges
The survey identifies challenges in evaluating financial graphs, explaining GNNs, formulating tasks, reproducing studies, and scaling models to large datasets.
- Graph evaluation methods: Graph comparisons are often inadequate, especially when similarity thresholds produce substantially different graphs and affect model performance.The survey emphasizes evaluating graph quality and justifying threshold choices, including through automated threshold selection.
- Explainability: Financial GNN explainability remains limited for heterogeneous and dynamic graphs, despite its importance for decision-making and reducing economic losses.Existing work focuses mainly on simpler graphs; edge-attributed and dynamic graphs remain underexplored.
- Task type: Financial GNN applications mostly use node-level tasks, leaving comparatively rare edge-level and graph-level formulations as opportunities for future work.The survey notes that supply-chain mining is one example of link prediction in finance.
- Data availability: Only about 24% of reviewed papers release code, making reproduction and methodological comparison difficult and motivating shared benchmark datasets.Financial data may also be restricted by corporate privacy obligations.
- Scalability: Commercial financial graphs can be extremely large, making scalability difficult because Laplacian computation and optimization are challenging on irregular graphs.Sampling can partially reduce the burden but may lose structural information.
Supplementary material
The supplementary materials provide application summaries, graph-type methodology figures, and the acronyms used throughout the text.
- Supplementary materials contain summary tables for each financial application, figures categorizing major GNN methodologies by graph type, and text acronyms.
- The supplementary materials extend the main text with application summaries, methodology classifications, and acronym references.
- The supplementary materials include materials not covered in the main text.
Figures on GNN methods for each graph type
The figures organize GNN pipelines by graph type, showing how convolution, aggregation, splitting, and prediction components differ across homogeneous, bipartite, and multi-relation graphs.
- Homogeneous graphs: For homogeneous graphs, adjacency and feature matrices enter graph convolution, neighbor information is aggregated, and an MLP produces node labels.
- Bipartite graphs: For bipartite graphs, convolution updates same-type node representations using self and neighbor information before aggregation across both node types and MLP prediction.
- Multi-relation graphs: For multi-relation graphs, edges are split into same-type subgraphs, convolution operates within each subgraph, and between-graph aggregation yields final node representations.
- Notation: The figure defines Conv as graph convolution, MLP as multi-layer perception, Agg as aggregation, and Split as characteristic-based data splitting.
Acronyms
The acronym list expands methods, models, datasets, financial tasks, evaluation metrics, and technical terms used throughout the survey.
- Models and forecasting: The list includes financial prediction, ranking, and forecasting models such as ARIMA, LSTM, RSR, STAR, and TRACER.
- Metrics: Evaluation and statistical terms include AUC, MAE, MCC, MRR, MSE, RMSE, Precision@k, KS, and IRR.
- Graph methods: Graph-learning methods include GCN, GAT, RGCN, HAN, HGT, GraphSAGE, and heterogeneous or bipartite variants.
- Applications: The glossary covers fraud, anomaly, recommendation, and embedding methods including CARE-GNN, DOMINANT, RCF, GEM, and LINE.