Identifying Evidence Subgraphs for Financial Risk Detection via Graph Counterfactual and Factual Reasoning

TL;DR

CF3 tackles the interpretability challenge in financial risk detection on company knowledge graphs by jointly modeling structure and node features through counterfactual and factual reasoning. It introduces a meta-path based Granger-causality attribution, an edge-type graph generator, and a layer-based feature masker to extract evidence subgraphs that causally drive predictions. The method is trained with a composite loss combining prediction, reconstruction, and reasoning objectives, and evaluated on three real-world bankruptcy datasets where it outperforms 13 baselines. The approach yields faithful, global explanations that can be produced in inductive settings, supporting timely risk warning and systemic risk mitigation.

Abstract

Company financial risks pose a significant threat to personal wealth and national economic stability, stimulating increasing attention towards the development of efficient andtimely methods for monitoring them. Current approaches tend to use graph neural networks (GNNs) to model the momentum spillover effect of risks. However, due to the black-box nature of GNNs, these methods leave much to be improved for precise and reliable explanations towards company risks. In this paper, we propose CF3, a novel Counterfactual and Factual learning method for company Financial risk detection, which generates evidence subgraphs on company knowledge graphs to reliably detect and explain company financial risks. Specifically, we first propose a meta-path attribution process based on Granger causality, selecting the meta-paths most relevant to the target node labels to construct an attribution subgraph. Subsequently, we propose anedge-type-aware graph generator to identify important edges, and we also devise a layer-based feature masker to recognize crucial node features. Finally, we utilize counterfactual-factual reasoning and a loss function based on attribution subgraphs to jointly guide the learning of the graph generator and feature masker. Extensive experiments on three real-world datasets demonstrate the superior performance of our method compared to state-of-the-art approaches in the field of financial risk detection.

Figures (5)

• Figure 1: An example of extracting explanations for company financial risk detection. Based on the computational subgraph of target company node A, (a), (b), and (c) are the explanations extracted by counterfactual reasoning, factual reasoning, and CF3, respectively. The subgraph in (c) also represents the ground-truth explanation.
• Figure 2: The overview of our proposed method: CF3. (a) Meta-path based attribution process: The aim is to generate the attribution subgraph for the target node by sequentially applying masks to each meta-path. (b) Graph generator: Designed with an encoder-decoder architecture, this module outputs reconstructed subgraph structures for the target node under different relationship networks. (c) Layer-based feature mask: Intended to produce feature mask matrices for different GCN layers and a global feature mask matrix. (d) Counterfactual and factual reasoning: The objective is to infer important subgraph structures and crucial node features through reasoning based on factual and counterfactual scenarios. The target GNN (such as HAT zheng2021heterogeneous, DANSMP zhao2022stock, etc., used in the experimental section of this paper) is pre-trained, and the parameters would not be changed during the training of CF3. Please note that once training is complete, CF3 can utilize the (b) and (c) to construct explanations for the target GNN with minimal time consumption.
• Figure 3: (a)-(b) present the ablation analysis of CF3 on the SMEsD datasets, (c) on SME, and (d) shows the parameter analysis.
• Figure 4: The influence of different numbers of edge types and edges were selected for interpretation on SMEsD.
• Figure 5: Structures of the real subgraphs and evidence subgraphs. And the dotted lines denote the removed edges.

Theorems & Definitions (3)

• Definition 1
• Definition 2
• Definition 3