EVINET: Towards Open-World Graph Learning via Evidential Reasoning Network (2506.07288v2)

Abstract: Graph learning has been crucial to many real-world tasks, but they are often studied with a closed-world assumption, with all possible labels of data known a priori. To enable effective graph learning in an open and noisy environment, it is critical to inform the model users when the model makes a wrong prediction to in-distribution data of a known class, i.e., misclassification detection or when the model encounters out-of-distribution from novel classes, i.e., out-of-distribution detection. This paper introduces Evidential Reasoning Network (EVINET), a framework that addresses these two challenges by integrating Beta embedding within a subjective logic framework. EVINET includes two key modules: Dissonance Reasoning for misclassification detection and Vacuity Reasoning for out-of-distribution detection. Extensive experiments demonstrate that EVINET outperforms state-of-the-art methods across multiple metrics in the tasks of in-distribution classification, misclassification detection, and out-of-distribution detection. EVINET demonstrates the necessity of uncertainty estimation and logical reasoning for misclassification detection and out-of-distribution detection and paves the way for open-world graph learning. Our code and data are available at https://github.com/SSSKJ/EviNET.

• The paper introduces EviNet, an evidential reasoning network that leverages Beta embeddings for robust misclassification and OOD detection.
• It utilizes Dissonance and Vacuity Reasoning modules to estimate uncertainty and logically differentiate known classes from novel data.
• Empirical evaluations on five benchmarks demonstrate state-of-the-art performance in classification accuracy and uncertainty estimation metrics.

EviNet: Evidential Reasoning Network for Resilient Graph Learning in Open and Noisy Environments

The presented paper introduces EviNet, an Evidential Reasoning Network designed to enhance graph learning applications in open and noisy environments. Traditional graph learning methods are often constrained by a closed-world assumption whereby all possible data labels are known a priori. This assumption, however, does not hold true in settings characterized by data ambiguity and novelty, such as financial fraud detection and other real-world scenarios. EviNet addresses two critical challenges in this context: (1) misclassification detection for in-distribution data, and (2) out-of-distribution (OOD) detection for data from novel classes.

The core architecture of EviNet integrates Beta embedding into a subjective logic framework, focusing specifically on two key components: Dissonance Reasoning and Vacuity Reasoning. The former is tailored for detecting misclassifications, while the latter is designed for OOD data detection. Together, they facilitate accurate uncertainty estimation and logical reasoning, paving the way for deployment in open-world graph learning tasks.

Model Structure and Methodology

EviNet's architecture leverages Beta embedding, a strategy introduced to endow nodes within the graph with the capacity for logical reasoning. The encoding of nodes in this probabilistic framework is advantageous for capturing uncertainty, a pivotal aspect when evaluating the likelihood of class membership under ambiguous conditions. This framework's disjunction and negation operations are central to deriving support regions for both known and novel classes, enhancing EviNet's proficiency in OOD detection tasks.

1. Dissonance Reasoning: This module employs logical operations to generate class embeddings for known categories based on training data, enabling the model to quantify conflicting evidence. The module effectively measures the dissonance score, which signifies the probability of a node being misclassified into one of the known classes.
2. Vacuity Reasoning: Building upon the disjunction operation results, Vacuity Reasoning characterizes novel or OOD classes by defining their implicit support region through logical negation. This module assesses the vacuity score, indicating the likelihood of a node belonging to novel classes absent from the training dataset.

EviNet dynamically computes context-aware embeddings by constructing class-specific graphs and integrating these into the reasoning modules. This capability enhances the detection and classification accuracy observed in complex environments where data exhibits variety and noise.

Empirical Evaluation and Results

The paper clearly demonstrates EviNet's superiority through extensive experimentation on five benchmark datasets, including Amazon-Photo and Coauthor-CS, among others. When evaluated on tasks of in-distribution classification, misclassification detection, and OOD detection, EviNet consistently outperforms state-of-the-art baselines, achieving the following:

• In-distribution Classification: EviNet maintains parity with, and often surpasses, existing methods in terms of classification accuracy across datasets.
• Misclassification Detection: EviNet outperforms competitors, evidenced by achieving the lowest Area Under the Risk Coverage Curve (AURC) on all evaluated datasets.
• Out-of-distribution Detection: The model consistently delivers top performance for metrics like False Positive Rate at 95% True Positive Rate (FPR95) and the Area Under the Receiver Operating Characteristic (AUROC) curve, indicative of robust OOD detection capabilities.

These results not only validate EviNet's architecture and methodology but also affirm the efficacy of employing a structured reasoning framework augmented by Beta embeddings for uncertainty detection in graph learning tasks.

Implications and Future Directions

The introduction of EviNet provides a significant advancement in machine learning models applied to graphs, especially within environments characterized by a high level of ambiguity and novelty. The dual reasoning strategy inherently balances the conflicting requirements of accurate classification and reliable uncertainty estimation. This innovation has profound implications for high-stakes applications, such as financial fraud detection and other scenarios involving dynamic and evolving datasets.

For future research, expanding EviNet to accommodate larger and more varied graph structures could further consolidate its applicability. Additionally, integration with other graph neural network architectures might enhance scalability and applicability across diverse domains and further explore speculative improvements in real-time and online learning scenarios. This paper represents a meaningful step forward in the evolution of graph learning techniques equipped to handle the uncertainties inherent in open-world environments.