Data

Tools

Indexes

Applications

Experiments

Open Source Science

icml19-egocnn

Code for "Distributed, Egocentric Representations of Graphs for Detecting Critical Structures" (ICML 2019)

https://github.com/rctzeng/icml19-egocnn

Science Score: 44.0%

This score indicates how likely this project is to be science-related based on various indicators:

  • ✓
    CITATION.cff file
    Found CITATION.cff file
  • ✓
    codemeta.json file
    Found codemeta.json file
  • ✓
    .zenodo.json file
    Found .zenodo.json file
  • ○
    DOI references
  • ○
    Academic publication links
  • ○
    Academic email domains
  • ○
    Institutional organization owner
  • ○
    JOSS paper metadata
  • ○
    Scientific vocabulary similarity
    Low similarity (6.0%) to scientific vocabulary

Keywords

convolutional-neural-networks graph-convolutional-neural-networks graph-embeddings graph-networks graph-neural-networks icml icml-2019 interpretability scale-free-networks self-similarity
Last synced: 6 months ago · JSON representation ·

Repository

Code for "Distributed, Egocentric Representations of Graphs for Detecting Critical Structures" (ICML 2019)

Basic Info
Statistics
  • Stars: 20
  • Watchers: 1
  • Forks: 4
  • Open Issues: 0
  • Releases: 0
Topics
convolutional-neural-networks graph-convolutional-neural-networks graph-embeddings graph-networks graph-neural-networks icml icml-2019 interpretability scale-free-networks self-similarity
Created almost 7 years ago · Last pushed over 4 years ago
Metadata Files
Readme License Citation

README.md

Ego-CNN

This is the repo for "Distributed, Egocentric Representations of Graphs for Detecting Critical Structures", Ruo-Chun Tzeng, Shan-Hung Wu, In Proceedings of ICML 2019. * slides

In the paper, we proposed Ego-Convolution layer, which keeps the nice properties of Convolution layer to the graph including: * detection of location-invariant patterns * enlarged receptive fields in multi-layer architecture * [most importantly] detection of precise patterns

This enables our Ego-CNN to provide explanation to its prediction when jointly learned with a task. picture 1. In effect, Ego-CNN with L layers can detect patterns up-to L-hop ego-networks. 2. By using the existing CNN visualization techniques such as Transposed Convolution or Grad-CAM variants, we can visualize the detected patterns in a specific filter or a specific neuron. 3. By tying the weight of filter across different layers, our Ego-CNN is regularized to detect self-similar patterns

Dependence

  • Python >= 3.6
  • Tensorflow >= 1.0
  • NetworkX 2.0
  • Numpy >= 1.13, Matplotlib >= 2.1
  • Optparse

To Reproduce Our Result On ICML'19

Step 1. Download and Preprocess Graph Classification Datasets

Execute Command python download_dataset.py to download all the bioinformatic and social network datasets used in the paper.

Step 2. Train Ego-CNN on specified datasets for specified tasks

To reproduce ... * Graph Classification Experiments: run ./execute-graph-classification-on-benchmarks.sh * Effectiveness of Scale-Free Regularizer: run ./execute-graph-classification-on-benchmarks.sh * Visualization on synthetic compounds: run ./execute-graph-classification-on-benchmarks.sh

Owner

  • Name: Ruochun Tzeng
  • Login: rctzeng
  • Kind: user
  • Location: Sweden

KTH PhD in graph mining.

Citation (CITATION.cff) 

# YAML 1.2
---
abstract: "This is the repository for Distributed, Egocentric Representations of Graphs for Detecting Critical Structures (ICML 2019)."
authors: 
  -
    family-names: Tzeng
    given-names: "Ruo-Chun"
    orcid: "https://orcid.org/0000-0002-4222-274X"
cff-version: "1.1.0"
date-released: 2019-06-22
license: MIT
message: "If you use this software, please cite it using these metadata."
repository-code: "https://github.com/rctzeng/EgoCNN"
title: "Ego-CNN: Distributed, Egocentric Representations of Graphs for Detecting Critical Structures"
...

GitHub Events

Total
Last Year