The official implementation for paper "EVINET: Towards Open-World Graph Learning via Evidential Reasoning Network"

This repository contains the code for Evidential Reasoning Network(EviNet), a framework designed to address graph learning challenges in open and noisy environments. Traditional graph models struggle handling both misclassified samples and out-of-distribution (OOD) samples. EviNet leverages Beta embeddings and subjective logic to estimate uncertainty and effectively detect both types of anomalies.

1. Dissonance Reasoning (M1): Detects misclassifications by computing a dissonance score that measures conflicting evidence, indicating samples that belong to multiple known classes.
2. Vacuity Reasoning (M2): Identifies OOD data by detecting a lack of evidence for known classes, using logical negation to create implicit support regions for novel classes.

EviNet achieves robust performance by balancing in-distribution classification, misclassification detection, and OOD detection. The framework provides support regions for known and novel classes via logical reasoning, helping users handle uncertainties in real-world scenarios.

1. Clone this repository.

2. Install the dependencies to conda via env.yaml.

For reproducing the baselines' results in Table 1

cd baseline
bash run_base.sh

For reproducing EviNet results in Table 1

bash run.sh

For reproducing Table 2 and Table 3

bash run_ablation.sh

The results can be found in under ./baseline/results/ for baseline and ./results/ for EviNet and ablation studies.

If you find the project helpful, we would appreciate if you cite the work:

@inproceedings{
  anonymous2025evinet,
  title={{EVINET}: Towards Open-World Graph Learning via Evidential Reasoning Network},
  author={Anonymous},
  booktitle={31st SIGKDD Conference on Knowledge Discovery and Data Mining - Research Track (February 2025 Deadline)},
  year={2025},
  url={https://openreview.net/forum?id=SCRDyuA6Ju}
}