arXivTimes / arXivTimes

repository to research & share the machine learning articles
https://arxivtimes.herokuapp.com/
MIT License
3.88k stars 202 forks source link

Semi-Supervised Classification with Graph Convolutional Networks #21

Open ryamauchi opened 7 years ago

ryamauchi commented 7 years ago

一言でいうと

与えられたグラフのノードを半教師ありクラスタリングする。既存の研究では隣接ノードは同じクラスであるという前提を正則化項として加えていたためモデルのキャパシティが制限されていたが(エッジのもつ情報はsimilarityとは限らない)、グラフ構造をニューラルネットとして表現することでその制約を取り払った。各Layerにおいて隣接行列が登場するので例えば3層ネットワークの場合は3段飛びの関係が考慮されることになり、それが正則化項の代わりをする(と思われる)。

論文リンク

https://arxiv.org/abs/1609.02907

著者/所属機関

Thomas N. Kipf, Max Welling

概要

https://tkipf.github.io/graph-convolutional-networks/

新規性・差分

手法

結果

コメント

icoxfog417 commented 6 years ago

PyTorch実装 https://github.com/tkipf/pygcn

TensorFlow実装 https://github.com/tkipf/gcn

icoxfog417 commented 5 years ago

解説スライド https://www.slideshare.net/DeepLearningJP2016/dl-hackssemisupervised-classification-with-graph-convolutional-networks-121489644