EviSEC: Evidential Spectrum-Aware Contrastive Learning for Out-of-Distribution Detection in Dynamic Graphs

🔥 Update

• [2024-5-26]: EviSEC is accepted by ECML 2025! (Acceptance Rate: 24%)
• [2025-6-10]: Paper of EviSEC online. Check out for details.

Abstract

In this study, we explore Out-of-Distribution Detection in Dynamic Graphs and analyze it using Evidential Deep Learning. We employ the 6 datasets in three download task and implement our methods for a comprehensive analysis of results. Specifically, we propose EviSEC, an innovative and effective OOD detector via Evidential Spectrum-awarE Contrastive Learning.

Dataset

6 datasets were used in the paper:

	# Nodes	# Edges	# Time Splits	Task
BC-OTC	5,881	35,588	95 / 14 / 28	Edge Classification
BC-Alpha	3,777	24,173	95 / 13 / 28	Edge Classification
UCI	1,899	59,835	62 / 9 / 17	Link Prediction
AS	6,474	13,895	70 / 10 / 20	Link Prediction
Elliptic	203,769	234,355	31 / 5 / 13	Node Classification
Brain	5,000	1,955,488	10 / 1 / 1	Node Classification

Data Source

Bitcoin OTC: Downloadable from http://snap.stanford.edu/data/soc-sign-bitcoin-otc.html

Bitcoin Alpha: Downloadable from http://snap.stanford.edu/data/soc-sign-bitcoin-alpha.html

Uc_irvine: Downloadable from http://konect.uni-koblenz.de/networks/opsahl-ucsocial

Autonomous Systems: Downloadable from http://snap.stanford.edu/data/as-733.html

Elliptic: Please see the instruction to manually prepare the preprocessed version or refer to the following repository that originally proposed the usage of the data: https://arxiv.org/abs/1902.10191

Brain: Downloadable from https://www.dropbox.com/sh/33p0gk4etgdjfvz/AACe2INXtp3N0u9xRdszq4vua?dl=0

For downloaded data sets please place them in the 'data' folder. Elliptic can also be easily processed by ell_preprocess.py.

Code Execution Method

To reproduce this study, the following code execution methods were used:

1. Conda Environment Requirements

• Python version: 3.6.13

• Dependencies:

  $ conda create --name <env> python=3.6.13 --file environment.txt

  $ pip install -r requestment.txt

2. Data Preprocessing

• The code performs data preprocessing, including data OOD (SM and FI).

3. Usage (Our model uses ----EDL evisec as part of the command line arguments.)

• python run_exp.py --config_file ./experiments/EC_BTCAlpha.yaml --OOD FI
  python run_exp.py --config_file ./experiments/EC_BTCAlpha.yaml --OOD SM
  python run_exp.py --config_file ./experiments/EC_BTCAlpha.yaml --EDL evisec --OOD FI
  python run_exp.py --config_file ./experiments/EC_BTCAlpha.yaml --EDL evisec --OOD SM