Zviri Mukati[Viga][Ratidza]
Maitiro ekudzidza akadzama anozivikanwa se "graph neural network" (GNNs) anoshanda mudura regirafu. Manetiweki aya achangobva kuwana kushandiswa munzvimbo dzakasiyana siyana, kusanganisira kuona komputa, masisitimu ekukurudzira, uye combinatorial optimization, kudoma mashoma.
Pamusoro pezvo, manetwork aya anogona kushandiswa kumiririra masisitimu akaomarara, anosanganisira masocial network, protein-protein interaction network, ruzivo magirafu, uye mamwe mundima dzinoverengeka dzekufunda.
Iyo isiri-euclidean nzvimbo ndipo panoshanda graph data, mukupesana nemamwe marudzi edata semifananidzo. Kuti ugoronga node, kufanotaura zvinongedzo, uye data remasumbu, kuongororwa kwegirafu kunoshandiswa.
Muchikamu chino, tichaongorora Grafu Neural Network zvakadzama, marudzi ayo, pamwe nekupa mienzaniso inoshanda uchishandisa PyTorch.
Saka, chii chinonzi Graph?
Girafu imhando yedata data inoumbwa nemanodhi uye vertices. Kubatana pakati pemanodhi akasiyana-siyana kunotemerwa nema vertices. Kana kutungamirirwa kwakaratidzwa mumanode, girafu inonzi inotungamirirwa; kana zvisina kudaro, hazvina kunanga.
Kushandiswa kwakanaka kwemagirafu kuenzanisira hukama pakati pevanhu vakasiyana-siyana mu pasocial network. Paunenge uchitarisana nemamiriro ezvinhu akaoma, akadai sekubatanidza uye kuchinjana, magirafu anobatsira zvikuru.
Ivo vanoshandirwa neanokurudzira masisitimu, semantic ongororo, social network kuongororwa, uye kucherechedzwa kwemaitiro
. Kugadzira girafu-yakavakirwa mhinduro ibasa-nyowani munda unopa kunzwisisa kunonzwisisika kwe data yakaoma uye yakabatana.
Grafu Neural Network
Graph neural network inyanzvi neural network mhando dzinogona kushanda pane graph data fomati. Girafu embedding uye convolutional neural network (CNNs) zvine zvazvinoita pazviri.
Graph Neural Networks inoshandiswa mumabasa anosanganisira kufanotaura node, mipendero, uye magirafu.
- CNN's inoshandiswa kurongedza mifananidzo. Saizvozvo, kufanotaura kirasi, maGNN anoiswa kune pixel grid inomiririra chimiro chegirafu.
- Kupatsanura zvinyorwa uchishandisa recurrence neural network. MaGNN anoshandiswawo nema graph architecture apo izwi rega rega mumutsara inzvimbo.
Kuitira kufanotaura node, mipendero, kana magirafu akazara, neural network inoshandiswa kugadzira maGNN. Kufanotaura padanho renode, semuenzaniso, kunogona kugadzirisa dambudziko sekuona spam.
Kufanotaura kwekubatanidza inyaya yakajairwa mumasisitimu ekurumbidza uye inogona kunge iri muenzaniso wedambudziko rekufanotaura.
Grafu Neural Network Mhando
Mazhinji neural network marudzi aripo, uye Convolutional Neural Networks aripo mune mazhinji acho. Tichadzidza nezve anonyanya kuzivikanwa maGNN muchikamu chino.
Graph Convolutional Networks (GCNs)
Iwo anofananidzwa neClassic CNNs. Inowana hunhu nekutarisa padyo node. Iyo activation basa rinoshandiswa neGNNs kuwedzera isiri-mutsara mushure mekubatanidza node mavheji uye kutumira zvinobuda kune dense layer.
Iyo inoumbwa neGrafu convolution, mutsara mutsara, uye isiri-yevadzidzi activation basa, mukukosha. MaGCN anouya mumhando mbiri huru: Spectral Convolutional Networks uye Spatial Convolutional Networks.
Grafu Auto-Encoder Networks
Inoshandisa encoder kudzidza kumiririra magirafu uye decoder kuyedza kugadzira patsva magirafu ekuisa. Pane bhodhoro remukati rinobatanidza encoder uye decoder.
Sezvo auto-encoders ichiita basa rakanakisa rekubata kirasi chiyero, ivo vanowanzo shandiswa mukubatanidza kufungidzira.
Recurrent Graph Neural Networks (RGNNs)
Mune akawanda-ane hukama network, uko imwe node ine hukama hwakawanda, inodzidza iyo yakakwana yekuparadzira maitiro uye inogona kubata magirafu. Kuti uwedzere kutsetseka uye kuderedza pamusoro-parameterization, zvinojairika zvinoshandiswa mune iyi fomu yegraph neural network.
Kuti uwane mhedzisiro iri nani, maRGNN anoda mashoma ekugadzirisa simba. Iwo anoshandiswa kugadzira zvinyorwa, kucherechedzwa kwekutaura, kududzira muchina, tsananguro yemifananidzo, kutagi kwevhidhiyo, uye kupfupisa zvinyorwa.
Gated Neural Graph Networks (GGNNs)
Kana zvasvika kumabasa enguva refu anotsamira, anodarika maRGNN. Nekusanganisira node, mupendero, uye magedhi epanguva pane kutsamira kwenguva refu, gated graph neural network inosimudzira inodzokororwa graph neural network.
Iwo masuwo anoshanda zvakafanana neGated Recurrent Units (GRUs) mukuti anoshandiswa kurangarira nekukanganwa data mumatanho akasiyana.
Kuita Graph Neural Network uchishandisa Pytorch
Iyo chaiyo nyaya yatichange tichitarisa pairi inyaya yakajairika node categorization. Tine social network yakakura inonzi musae-github, iyo yakagadzirwa kubva kuAPI yakavhurika, yeGitHub vanogadzira.
Mipendero inoratidza hukama hwevateveri pakati pemanodhi, ayo anomiririra vanogadzira (vashandisi vepuratifomu) vane nyeredzi mune angangoita gumi repositori (ona kuti izwi rekuti mutual rinoratidza hukama husina kutungamirwa).
Zvichienderana nenzvimbo yenzvimbo, marekodhi ane nyeredzi, mushandirwi, uye kero yeemail, node hunhu hunodzoserwa. Kufanotaura kana mushandisi weGitHub ari mugadziri wewebhu kana a mugadziri wekudzidza muchina ndiro basa redu.
Zita rebasa remushandisi wega wega rakashanda sehwaro hweiyi basa rekunongedza.
Kuisa PyTorch
Kutanga, tinofanira kutanga taisa PyTorch. Unogona kuzvigadzirisa zvinoenderana nemushini wako kubva pano. Heino yangu:
Kupinza mamodule
Iye zvino, isu tinopinza ma module anodiwa
Kupinza uye Ongorora iyo data
Nhanho inotevera ndeyekuverenga data uye kuronga mitsara mishanu yekutanga uye mitsara mishanu yekupedzisira kubva kune mavara faira.
Makoramu maviri chete pamakoramu mana — id yenodhi (kureva, mushandisi) uye ml_target, inova 1 kana mushandisi ari nhengo yenharaunda yekudzidza muchina uye 0 neimwe nzira-akakosha kwatiri mumamiriro ezvinhu aya.
Tichifunga kuti kune makirasi maviri chete, isu tinogona ikozvino kuve nechokwadi kuti basa redu inyaya yebhinary classification.
Nekuda kwekusaenzana kwekirasi kwakakosha, mugadziri anogona kungofungidzira kuti ndeipi kirasi ndiyo yakawanda pane kuongorora kirasi isina kumiririrwa, ichiita chiyero chekirasi chimwe chinhu chakakosha kufunga.
Kuronga histogram (frequency distribution) kunoburitsa kumwe kusaenzana nekuti kune mashoma makirasi kubva muchina kudzidza (label=1) pane kubva kune mamwe makirasi.
Feature Encoding
Manodhi 'maitiro anotizivisa nezve ficha inodyidzana neimwe node. Nekushandisa nzira yedu yekukodha data, isu tinogona ipapo kukodha iyo hunhu.
Tinoda kushandisa nzira iyi kuvharidzira chikamu chidiki chetiweki (titi, 60 nodes) yekuratidzira. Kodhi yakanyorwa pano.
Kugadzira uye kuratidza magirafu
Tichashandisa torch geometric. data kugadzira girafu yedu.
Kuenzanisira girafu rimwechete rine zvakasiyana (zvichida) zvivakwa, data iri nyore Python chinhu chinoshandiswa. Nekushandisa kirasi iyi uye hunhu hunotevera - ese ari ma torch tensor - isu tichagadzira yedu girafu chinhu.
Chimiro cheukoshi x, iyo ichagoverwa kune encoded node features, i[nhamba yemanodhi, nhamba yezvimiro].
Chimiro che y [nhamba yemanodhi], uye ichaiswa kune ma node label.
kumucheto indekisi: Kuti titsanangure girafu risina kurongeka, tinoda kuwedzera ma indices ekumucheto kuitira kuti tibvumire kuvapo kwemativi maviri akasiyana anonangidzirwa anobatanidza node mbiri dzakafanana asi anonongedza kumativi akasiyana.
Peya yemipendero, imwe inonongedza kubva ku100 kusvika ku200 uye imwe kubva ku200 kusvika ku100, inodiwa, semuenzaniso, pakati penode 100 uye 200. Kana iyo yekumucheto indices ichipihwa, zvino iyi ndiyo nzira iyo isina kurongeka girafu inogona kumiririrwa. [2,2* nhamba yemacheto ekutanga] ichava chimiro che tensor.
Isu tinogadzira nzira yedu yekudhirowa girafu kuratidza girafu. Danho rekutanga nderekushandura network yedu ine homogeneous kuita NetworkX graph, iyo inogona kubva yadhonzwa uchishandisa NetworkX.draw.
Gadzira yedu GNN modhi uye uidzidzise
Isu tinotanga nekukodha iyo yese seti yedata nekuita encode data nechiedza = Nhema uye todaidza kuvaka girafu nechiedza = Nhema kuvaka iyo girafu yese. Hatisi kuzoedza kudhirowa iri hombe girafu nekuti ndiri kufungidzira kuti uri kushandisa muchina wemuno une zviwanikwa zvishoma.
Masks, ari mabhinari mavheji anotaridza kuti ndeapi node dzemask yega yega uchishandisa manhamba 0 uye 1, anogona kushandiswa kuzivisa chikamu chekudzidzisa icho node dzinofanirwa kuverengerwa panguva yekudzidziswa uye kuudza chikamu chekufungidzira kuti ndedzipi dhata rebvunzo. Torch geometric.inoshandura.
Node-level split inogona kuwedzerwa uchishandisa mask yekudzidzira, val mask, uye test mask zvimiro zveAddTrainValTestMask kirasi, inogona kushandiswa kutora girafu uye kutigonesa kutsanangura magadzirirwo atinoda kuti masiki edu avakwe.
Isu tinongoshandisa gumi muzana pakudzidzisa uye kushandisa makumi matanhatu muzana e data seyedzo yakatarwa tichishandisa makumi matatu muzana seti yekusimbisa.
Ikozvino, isu tichaisa maviri GCNConv akaturikidzana, yekutanga ine inobuda ficha kuverenga iyo yakaenzana nehuwandu hwezvimiro mugirafu yedu seyekupinza maficha.
Muchikamu chechipiri, chine node dzinobuda dzakaenzana nenhamba yemakirasi edu, tinoshandisa relu activation basa uye tinopa iyo yakadzikama maficha.
Edge index uye uremu hwemupendero zviviri zvezvakawanda sarudzo x izvo GCNConv inogona kugamuchira mune yekumberi basa, asi mumamiriro edu ezvinhu, isu tinongoda maviri ekutanga akasiyana.
Kunyangwe chokwadi chekuti modhi yedu ichakwanisa kufanotaura kirasi yenode yega yega mugirafu, isu tichiri kuda kuona kurongeka uye kurasikirwa kweseti imwe neimwe zvakasiyana zvichienderana nechikamu.
Semuenzaniso, panguva yekudzidziswa, isu tinongoda kushandisa iyo yekudzidziswa seti kuona iko kurongeka uye kurasikirwa kwekudzidziswa, uye saka apa ndipo panouya masiki edu anoshanda.
Kuti tiverenge kurasikirwa kwakakodzera uye kurongeka, isu tichatsanangura mabasa ekurasikirwa kwemasiki uye kurongeka kwechokwadi.
Kudzidzisa muenzaniso
Zvino zvatatsanangura chinangwa chekudzidzisa icho torch ichashandiswa. Adam ndiye master optimizer.
Tichaitisa dzidziso yeimwe nhamba yenguva tichiramba takatarisa chokwadi chechokwadi.
Isu tinoronga zvakare kurasikirwa kwekudzidziswa uye hunyanzvi mukati menguva dzakasiyana dzakasiyana.
Zvakaipa zveGrafu Neural Network
Kushandisa maGNN kune mashoma mashoma. Irini rekushandisa GNNa uye maitiro ekusimudzira mashandiro emamodhi edu ekudzidza muchina zvese zvichajekeswa kwatiri mushure mekunge tanyatsonzwisisa.
- Nepo maGNN ari ma network asina kudzika, anowanzo aine matatu akaturikidzana, mazhinji neural network anogona kudzika zvakadzika kuti avandudze mashandiro. Hatikwanise kuita padanho rekucheka pamaseti makuru nekuda kwekupikiswa uku.
- Zvakanyanya kuoma kudzidzisa modhi pamagirafu, sezvo maitiro avo ekugadzirisa ane simba.
- Nekuda kwemitengo yakakwira yemakomputa emanetiweki aya, kuyera modhi yekugadzira kunopa matambudziko. Kuyera maGNN ekugadzira kuchave kwakaoma kana chimiro chako chegirafu chakakura uye chakaoma.
mhedziso
Kwemakore mashoma apfuura, maGNN akagadzira maturusi ane simba uye anoshanda emuchina wekudzidza nyaya mudura regirafu. Iyo yakakosha yekutarisa yegraph neural network inopiwa mune ino chinyorwa.
Mushure meizvozvo, unogona kutanga kugadzira iyo dataset iyo ichashandiswa kudzidzisa uye kuyedza modhi. Kuti unzwisise kuti inoshanda sei uye kuti inokwanisa sei, iwe unogona zvakare kuenda kure uye kuidzidzisa uchishandisa imwe mhando yedataset.
Happy Coding!
Leave a Reply