Ԫר
ߣ٩
༭ŷ
Ԫͼ(Graph NN)ǽһоȵ㣬DeepMind“Graph Networks”ųѧϰʵƪĻɬѶǼϯAIѧҡҽﴴʼ˵٩ʿ廪ʿڽŶӶGNNĻϣDeepMind“ͼ”塣
- 1 -
ع 2018 ѧϰĽչ20186 DeepMind Ŷӷ
“Relational inductive biases, deep learning, and graph networks”
һƪҪģҵ顣
ܶѧǵ˼·оа廪ѧïŶӡ201812£һƪĿ“Graph neural networks: A review of methods and applications”
20191£ʿڽŶӣҲдһƪƪĸȫ棬Ŀ“A Comprehensive Survey on Graph Neural Networks”
ʿڽŶGNNԴarxiv
DeepMind ŶӵƪģҵôҵĹעԭ
Դ AlphaGo սʤhԺDeepMind ҵ磬ΪѧϰҵŶӣDeepMind Ŷӷģܵͬձע
ԴDeepMind Ŷӷ [1] Ժã Github ϿԴǿϵͳĿƽ Graph Nets [4]
⣺ͿԴҪDzDZҵҪԭҪԭ⣬DeepMind Ŷоǣѧϰͼס
- 2 -
ͼ (Graph) ɵ (Node) ͱ (Edge) ɡ
ͼһҪѧģͣܶ⡣
Ʃǰѳе·ͼͼףÿվһ㣬ڵĵվ֮߾DZߣ㵽յ㣬ǿͨͼļ㣬㵽յ㣬ʱ̡˴ٵг·ߡ
Ʃ Google Ͱٶȵ棬ÿվÿҳͼеһ㡣ÿҳӣվҳÿӶͼеһߡĸҳõԽ࣬˵ҳԽףǣҲԽǰ
ͼIJȻд
Ʃ뼸ε˾н·ߣÿн·ǰʱеһʱ䡢GPSγȣ顣ΰѼн·ߣһ𣬹еͼ
ѵͼҲһͼףÿ·ڣһ㣬ڵ·ڣһߡ
òƺܼϸں鷳
ٸӣ·кܶʽʮ·ڣdzڣлε——δӶ·Уȷ·ڵλã
ձ챣ɽţȷŵλ
- 3 -
ѧϰͼףܹǶͼĴ
ѧϰͼıĴ棬Ѿȡ˾ijɹѧϰijɹʹ֮Ӧͼ״
ͼɺƽֱؾɡһǶȣÿΪͼеһ㣬ÿصܱߵ 8 ֮䶼бߣÿ߶ȳͨӽǣͼͼǹͼһ
ͼѧϰֶΣԸͷ棬ӦڹͼףƩ convolutionresidualdropoutpoolingattentionencoder-decoder ȵȡѧϰͼ״뷨ʵܼ
Ȼ뷨ܼ뵽ϸڣսÿսζŸǿļŸDZӦó
ѧϰͼ״оҵûͳһijν
ǿͼѧԵŶӣоΪ Geometric Deep LearningïŶӺʿŶӣǿͼ״еҪԣǿ˼ԴǰΪ Graph Neural NetworksDeepMind ŶȴضֶΣʹøƣGraph Networks
ôҪַȥչȴҪѸѧɵĿ궨λͼڼǿ֮ͬ⣬ڴٽ֮ͬδ
- 4 -
ʿŶӰѧϰͼ״չ 5 ӷdzö
ʿŶӰѧϰͼ״ 5 ӷԴ A Comprehensive Survey on Graph Neural Networks
Graph Convolution Networks
Graph Attention Networks
Graph Embedding
Graph Generative Networks
Graph Spatial-temporal Networks
˵ Graph Convolution Networks (GCNs)
GCN ܣԴ A Comprehensive Survey on Graph Neural Networks
GCN CNN ӦڹͼסCNN ҪΪĸ
֮ںϡͼ֮ںҪͨ (convolution) ʵ֡ڹͼ֮ĹϵñԣڹͼںϣбȾǿİ취Messsage passing [5] һָǿİ취
ֲCNN ʹ convolution İ취ԭʼؾУ߲ĵ㣬ǹĵ㣬ںԡںڵİ취ҲӦڹͼС
CNN ʹ pooling ֶΣԭʼУԵڱԵУʵʵУ߲ʵ塣CNN ͨ convolution pooling ʹãṹӣܸǿ硣ڹͼףҲڻ Messsage passing Poolingͼס
㡣CNN ͨʹ softmax ֶΣͼз࣬ʶͼںڹͼ˵ĽԶͼףȵȽҲԤͼijضĵֵҲԤijߵֵ
GCN Graph Attention Networks Դ A Comprehensive Survey on Graph Neural Networks
Graph Attention Networks Ҫ⣬ GCN ƣڵںϡķ
Graph Convolution Networks ʹþʽʵֵںϺͷֲConvolution ʽںڵĵ㣬 attention ۽ʽȴڵĵ㣬ÿںͼ㣬ǷڣǷںںϣȡڵ֮Ĺǿ
Attention ǿǶҪߣΪҪͼ֮Ĺǿ Graph Attention Networks оص㣬νͼɱͨм㣬Чʡ
- 5 -
Graph Embedding Ҫ⣬Ǹͼÿÿߣһֵͼ⣬Ϊֵǣıִʻ乹ɣҪִʻ㣬תֵʹѧϰ㷨
ıеÿֻʻ㣬ͼеһ㣬ͬʱѴ֮ϵͼеһߣôͶ䣬͵ͬıͼеһн·
ܹÿֺʹʻ㣬һеֵôͶӦн··
жʵ Graph Embedding İ취ЧȽϺõİ취 Autoencoder GCN İ취ͼĵͱתֵ̳Ϊ (encoding)Ȼ֮ͨľ룬ֵϣתΪͼף̳Ϊ (decoding)ͨϵصΣýõͼףԽԽԭʼͼף̳Ϊѵ
Graph Embedding ͼеÿÿߣеֵͼĽṹ⡣
ͼн·δЩн·УʶЩЩ֮ߣѶȸǣûн·ѵͼľֲԼ֮ӦͼԣΰѾֲƴӳͼȫòЩ Graph Generative Networks Ҫ⡣
Graph Generative Networks ȽDZʵַʹ Generative Adversarial Networks (GAN)
GAN (generator) ͱ (discriminator) ֹɣ1. ѵУƩ纣н·²ݱͼӦóʲô2. ɳͼףαһн·3. Ӵα·ʵ·Уѡ·ñʶļ·αġ
ɵɵֲ˭α·˭ʵ·˵ɳͼףܽӽʵͼס
GCN 4 ͼ磬Դ A Comprehensive Survey on Graph Neural Networks
- 6 -
Ծ̬ͼ⣬ͼʱǶ̬ģƩͼбֵĵ·Ǿ̬ģ·Ƕ̬ġ
Ԥⴺڼ䣬찲ŸĽͨӵ⣬Ҫǿռ spatial أƩ찲ܱߵĵ·ṹҲҪʱ temporal أƩ괺ڼõͨӵ Graph Spatial-temporal Networks Ҫ֮һ
Graph Spatial-temporal Networks ܽ⣬ƩһƵÿһ֡ͼУʶλãѵڣƵijЩ֡УпǿģƩ类Աڵˡ
ʱͨ˼· RNN LSTM GRU ȵȡ
DeepMind Ŷ RNN ϣ˱ͽ (encoder-decoder) ơ
- 7 -
DeepMind Ŷӵƪ[1]ԼĹ“part position paper, part review, and part unification”ںϡ⻰ô⣿
DeepMindϹȸԡMITȻ27λ߷ذģ“ͼ”Graph network˵ѧϰϣѧϰйϵ⡣
ǰ˵ʿŶӰѧϰͼ״չ 5 ӷ1) Graph Convolution Networks2) Graph Attention Networks3) Graph Embedding4) Graph Generative Networks5) Graph Spatial-temporal Networks
DeepMind Ŷ 5 ӷ 4 ֱ Graph Attention NetworksGraph EmbeddingGraph Generative Networks Graph Spatial-temporal Networksǰĸijɹ“ں”ͳһĿܣΪ Graph Networks
ǵУĸӷ;ɹ“”Dzû Graph Convolution Networks ijɹȻǴĸӷɹУѡΪDZķγԼ“”ǿԴĴ [4]
DeepMind201810¿ԴGraph Nets libraryTensorFlowйǿĹϵ硣Դgithub.com/deepmind/graph_nets
ȻУǵĸӷ⣬Dz鿴ǿԴĴ룬ʵǺӷGraph Attention Networks Graph Spatial-temporal Networks
DeepMind ˼·ģȣ [5] message passing ںϵĻƣ [6] ͼȫֵľ۽ϣͨõ graph block ģ飻Σ LSTM Ҫڽ encoder-decoder ܣʱлƣ graph block ģڽ encoder-decoder ܣγ Graph Spatial-temporal Networks ͨϵͳ
- 8 -
Ϊʲô DeepMind ijɹҪ¹ļ¡
һѧϰ̵Ľ
ԭϽѧϰƩ CNN ijɹڶͼIJϳҲǣԭʼؾУ߶Ρβ߶Уʵʵ壬ʵ塣
ǣ̽ CNN ÿһмʵϺȷһЩڵ㣬Ҳ֪һЩڵ㣬ʵ塣֮ܶCNN ṹǸԣȷؽṹصĹ̵ϸڡ
̵ϸڣҲ̸ΪԤ CNN ⣬ֻѵѵĽǷܴﵽڴЧϡº«ưȱݣȴȱݡ
˵ȷظ CNN ̵ϸڣͿԵصεĸڵIJΪԤ
Сѧϰ
ѧϰѵݣѵݵĹģܴͨռôѵݣҪ֯ȥѵݽбעǾս
ѧϰĹϸڣи˽⣬ǾͿԸƾøٵѵݣѵɵѧϰģ͡
Ḷ́Ḷ́ڵĵ㣬һ©زؽо
ǶԵ֮Ĺϵиȷ˽⣬ͲҪڵĵ㣬һ©زؽоֻҪйĵ㣬о
ݵ֮Ĺϵ磬ǹͼסͼĽṹͨ CNN ӼԣҪѵҲ١
Ǩѧϰ
õ CNNԴӴͼƬУʶijʵ壬Ʃè
ǣʶè CNN ʶèʶҪʶѵݡǨѧϰĹ̡
ܲܲṩʶѵݣֻùķʽߵè빷ȻõʶĿꡣ
ѧϰи˽⣬֪ܰʶڽѧϰ
ӹ巶Χ˵ѧϰ֪ʶͼףǻѧϰӪѧɵѧɡΪֹѧɸʤںѧɣȡ̣ѧܾõ⡣ѧϰ쵽ͼ״ѧɵںϣϣ
ġռʱںϣں
Ƶ˵ѧϰ߾硣
ƵںͼĿռָͼʵʶʵӦ⡣
̬֡ͼһγƵʵʱСͬһʵ壬ڲͬ֡λã̺ʵ˶˶ıɺ
δһƵܽı⡣߷θһı⣬ҵеƵƵľҲѶȳ
ο
Relational inductive biases, deep learning, and graph networkshttps://arxiv.org/abs/1806.01261
Graph neural networks: A review ofmethods and applicationshttps://arxiv.org/abs/1812.08434
A Comprehensive Survey on Graph Neural Networkshttps://arxiv.org/abs/1901.00596
Graph netshttps://github.com/deepmind/graph_nets
Neural message passing for quantum chemistryhttps://arxiv.org/abs/1704.01212
Non-local neural networkshttps://arxiv.org/abs/1711.07971