
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 Ababhali: 
 
 
 
 
 
 
 
 Zhen Peng Wenbing Huang Minnan Luo Qinghua Zheng Yu Rong Tingyang Xu Junzhou Huang ISISHWESCO Ubude be "content" kwii "networks" ezahlukeneyo ezifana nezi "social networks" kunye nezi "communication networks" zinika amathuba angazange abekho ngaphambili okokufunda "representations" ezisemgangathweni ngaphandle kwesikhokelo. Eli phepha ligxila kwindlela yokugcina nokukhupha ulwazi olubalulekileyo kwidatha ye "graph-structured" kwisithuba se "embedding" ngendlela engakho ikhokhelwa. Ngenxa yoku, siphakamisa ingcamango entsha, i "Graphical Mutual Information" (GMI), yokulinganisa ubudlelwane phakathi kwe "input graphs" kunye ne "high-level hidden representations". I "GMI" iphuhlisa ingcamango ye "mutual information" kwizikhompyutha ezisetyenziswa rhoqo kwisithuba se "vector" ukuya kwisithuba se "graph" apho ukulinganisa i "mutual information" ngeendlela ezimbini ze "node features" kunye "topological structure" kubalulekile. I "GMI" ibonisa izibonelelo ezininzi: Okokuqala, ayichaphazeleki yindlela yokujikeleza kwe "input graphs"—isithintelo esingenakuphetshwa kwiindlela ezininzi ze "graph representation learning" ezikhoyo; Ngaphandle koko, ingalinganiswa ngokufanelekileyo kwaye yongezwe ngeendlela zokulinganisa i "mutual information" zangoku ezifana ne "MINE"; Okokugqibela, uhlalutyo lwethu lwethiyori luqinisekisa ukuchaneka kwayo kunye nokulunga kwayo. Ngoncedo lwe "GMI", siyakha imodeli yokufunda engakho ikhokhelwa iqeqeshwe ngokwandisa i "GMI" phakathi kwe "input" kunye ne "output" ye "graph neural encoder". Uvavanyo olubalulekileyo kwi "transductive" kunye ne "inductive node classification" kunye ne "link prediction" lubonisa ukuba indlela yethu ingcono kunezinye iindlela ezingekho zikhokhelwa, kwaye ngamanye amaxesha ingcono kunezo zisebenzisa ikhokelo. 1      ISIVULA Ukufumana uhlalutyo oluthembekileyo lwe "quantitative analysis" kwi "information networks" ezahlukeneyo, ukufunda "representation learning" okusemgangathweni kwidatha ye "graph-structured" kuye kwaba ngomnye weemfundla ezibangela umdla ngoku. Iindlela zamva nje zokufezekisa le njongo, ezihlala ziqhelaniswa ne "Graph Neural Networks" (GNNs), ziye zenze inkqubela phambili emangalisayo kwimisebenzi emininzi yokufunda, njenge "node classification" [ ,  ,  ], "link prediction" [ ,  ,  ], "graph alignment" [ ,  ,  ], ukwenziwa kwe "molecular generation" [ ,  ,  ], ukubala nje ezimbalwa. Nangona inkqubela phambili inempumelelo, ukuqeqesha ii "GNNs" kwiindlela ezikhoyo ngokuqhelekileyo kufuna uhlobo oluthile lwe "supervision". Ngaphandle kwamathandabuzo, ulwazi lwe "label" lunzima ukulufumana—ukubhalwa ngesandla okanye ukuhlawula imali yemvume, kwaye kunokwenzeka ukuba kungafumaneki ngenxa yomgaqo-nkqubo wobumfihlo. Singatsho nje ukuba ukuthembeka kwezi "label" ezikhoyo ngamanye amaxesha kuyathandabuzeka. Ngoko ke, indlela yokufumana i "graph representation" esemgangathweni ngaphandle kwe "supervision" iyimfuneko kwimisebenzi emininzi esebenzayo, nto leyo enika inkuthazo yophando lweli phepha. 35 42 50 26 27 49 12 19 43 4 5 46 Ngokuhlolisiswa ngononophelo kwidatha yethu ye "graph-structured" eqokelelweyo, sifumanisa ukuba idla ngokubakho kwiimithombo ezahlukeneyo ezifana ne "social networks", i "citation networks", kunye ne "communication networks" apho ubuninzi bolwazi kunye nolwazi lwe "linkage" lubakho. Umzekelo, idatha kwiindawo ezininzi zentlalo ezinje nge "Twitter", "Flickr", kunye ne "Facebook" ibandakanya i "features" zabasebenzisi,  ., iinkcukacha zobuqu, iitekisi, imifanekiso, i "IP", kunye ne "relations" zabo,  ., ukuthenga into efanayo, ukuba ngabahlobo. Le datha yobuninzi "content" yanele ukuxhasa imisebenzi yoku "mining" ezayo ngaphandle kwesikhokelo esongezelekileyo: ukuba izinto ezimbini zibonisa ukufana okugqithisileyo kwi "features", kukho ithuba eliphezulu lokuba zibe ne "link" phakathi kwazo ("link prediction"), kwaye kusenokwenzeka ukuba zibe kwisikhombo esifanayo ("classification"); ukuba izinto ezimbini zikhomba kwenye into efanayo, kusenokwenzeka ukuba zibe neempawu ezifanayo ("recommendation"). Ngale ndlela, ukugcina nokukhupha ulwazi oluninzi kangangoko kwi "information networks" kwisithuba se "embedding" kunceda ukufunda "representations" ezisemgangathweni ezibonisa ukusebenza okufunwayo kwimisebenzi yok "mining" ngaphandle koluhlobo lwe "supervision". I "unsupervised graph representation learning" lukhetho oluthandwa kakhulu kwiimeko ezininzi ngenxa yenkululeko ekuswelekeni kwe "labels", ngakumbi xa sijonge ukufumana inzuzo kwidatha enkulu engabhalwanga. umzekelo umzekelo Ukuyifumana ngokupheleleyo idatha ebalulekileyo kwii "graphs", kule nqaku, siqhuba i "graph embedding" esekelwe kwi "Mutual Information" (MI) maximization, siphefumlelwe yimpumelelo ye "Deep Info-Max" [ ] esebenza kwimifanekiso. Ukufumana i "representations" eziluncedo, i "Deep InfoMax" iqeqesha i "encoder" ukwandisa i "MI" phakathi kwe "inputs" zayo ( ., imifanekiso) kunye ne "outputs" zayo ( ., i "vectors" ezingagciniyo). Xa sijonga i "Deep InfoMax" kwisithuba se "graph", ityala lokuqala esilifunayo ukulinyathela yindlela yokuchaza i "MI" phakathi kwama "graphs" kunye nee "vectors" ezingagciniyo, ngelixa i "topology" yama "graphs" ingaphezu kwe yeemifanekiso (bona uMfanekiso   Enye yeemicombekiso kukuqinisekisa ukuba umsebenzi we "MI" phakathi kwe "hidden representation" yomntu ngamnye kunye nee "features" zommelwane bayo uyayithobela imigaqo "symmetric property", okanye ngokulinganayo, ungachaphazeleki ngumbutho weendawo zommelwane. Njengomsebenzi wamva nje ojonga i "MI", i "Deep Graph Infomax" (DGI) [ ] yokuqala i "embeds" igraph yangena kunye negraph elahlekileyo ehambelanayo, emva koko ichaza igraph yangena njengevector ngokusebenzisa umsebenzi wokufunda, ekugqibeleni yandisa i "MI" phakathi kwe "summary vector" kunye nee "hidden representations" ngokukhetha igraph yangena (isampulu etyenzayo) kwigraph elahlekileyo (isampulu engemiyo). Ufumfana   ubonelela ngombono ongabambekiyo we "DGI". Ukwandisa olu hlobo lwe "MI" kuvele kube kukwandisa i "MI" phakathi kwe "node features" ezangenayo kunye nee "hidden vectors", kodwa eli "equivalence" libamba phantsi kweemeko ezithile ezibekiweyo,  ., umsebenzi wokufunda kufuneka ube "injective", nto leyo ebonakala ngathi iyalucinezela kakhulu kwiimeko zokwenyani. Nokuba singaqinisekisa ubukho bomsebenzi wokufunda "injective" ngeendlela ezithile zoyilo,  ., lowo usetyenziswa kwi "DeepSets" [ ], ubuchule bokufunda obunobuchule buyachaphazeleka kunye nendlela ye "parameters" yayo eyaqeqeshwa ngayo. Oko kukuthi, umsebenzi onobuchule wokufunda usenomngcipheko wokuba ungabi nobuchule ukuba uqeqeshwe ngaphandle kwe "supervision" yangaphandle. Kwaye ukuba umsebenzi wokufunda awunabuchule, ulwazi lwe "input graph" olukwi "summary vector" luya kuncipha njengoko ubungakanani begraph bungakanda. Ngaphezu koko, i "DGI" ihlala ikwi "MI maximization" yebubanzi obukhulu begraph/ipatch. Ngoko ke kwi "DGI", akukho siqinisekiso sokuba i "encoder" ingafumana ulwazi olwaneleyo kwidatha yangena njengoko ingadibanisi ngokucokisekileyo i "hidden representations" kunye nee "inputs" zazo zokwenyani. 20  Oko.  Oko. 1). 41 1 umzekelo umzekelo 47 Kule nqaku, siphakamisa indlela elula ngakumbi yokujonga i "MI" ngokubhekisele kwii "graphical structures" ngaphandle kokusebenzisa nawuphi na umsebenzi wokufunda okanye umsebenzi wokuchithachitha. Siyikhupha ngqo i "MI" ngokudibanisa i "input" ( ., i "sub-graph" equka iindawo zommelwane) kunye ne "output" ( ., i "hidden representation" yomntu ngamnye) ye "encoder". Kwaye ngokumangalisayo, ukuchazwa kwethu kwethiyori kubonisa ukuba i "MI" echazwe ngqo ingahlulwa ibe yingqokelela enomlinganiselo we "local MIS" phakathi kwe "neighbourhood features" nganye kunye ne "hidden vector". Ngale ndlela, sihlule i "input features" kwaye senze ukubala kwe "MI" kucace. Ngaphezu koko, olu hlobo lwe "MI" lunganelisekisa ngokulula imigaqo "symmetric property" ukuba silungisa amaxabiso ezilinganiselo. Siyibeka kude into eninzi kwi §   Njengoko i "MI" engentla igxile ikakhulu kubuninzi bez "node features", siyibiza ngokuba yi "Feature Mutual Information" (FMI).  Oko.  Oko. 3.1. Imiba emibini esele ikho malunga ne "FMI":   i "combining weights" ayikaziwa kwaye   ayibandakanyi i "topology" ( ., i "edge features"). Ukujongana nale micximba mibini, sichaza i "Graphical Mutual Information" (GMI) yethu esekelwe kwi "FMI". Ngokukhethekileyo, i "GMI" isebenzisa ixabiso elisemgangathweni ngokubeka i "weights" kwi "FMI" ilingana no "proximity" phakathi kommelwane ngamnye kunye nenqanawe ey target kwisithuba se "representation". Malunga nokugcina ulwazi lwe "topology", i "GMI" iyakwandisa la maxabiso kunye nee "edge features" zangena nge "mutual information term" eyongezelelweyo. I "GMI" eziphumelayo ayichaphazeleki yiyo i "topology" kwaye ikwangu "calculable" nge "Mutual Information Neural Estimation" (MINE)   Iimfuno eziphambili zomsebenzi wethu zezi zilandelayo: 1. 2.  Oko. . •    : Siphuhlisa i "MI estimation" yesintu kwisithuba se "graph" kwaye siphakamise ingcamango entsha ye "Graphical Mutual Information" (GMI) ngokufanelekileyo. I "GMI" ikhululekile kwi Iingcamango umngcipheko onokubakho obangelwa ngumsebenzi wokufunda njengoko ijonga i "MI" phakathi kwama "input graphs" kunye nee "high-level embeddings" ngendlela elula. •    : Ngohlalutyo lwethu lwethiyori, sinika ifom engabalekiyo nenokubalelwa ye "GMI" ehlula i Iialgorithm yonke i "GMI" ibe yingqokelela enomlinganiselo we "local MIS". Ngoncedo lwe "MINE method", i "GMI maximization" inokufumaneka ngokulula kwibubanzi benqanawe. •    : Siqinisekisa ukusebenza kwe "GMI" kwimisebenzi emininzi eyaziwayo yokuhlelwa kwe "node" kunye ne "link prediction" Izinto ezifunyenweyo ngovavanyo imiba kubandakanya i "transductive" kunye ne "inductive". Ukuvavanywa kubonisa ukuba indlela yethu ibonelela ngemisebenzi emihle kwii "benchmarks" ezahlukeneyo kwaye ngamanye amaxesha ingcono kunezo zisebenzisa ikhokelo. 2      IMIBUZO EPHATHELENE NOMSEBENZI Ngokuhambelana negxalaba lomsebenzi wethu, sibuyekeza ngokufutshane imisebenzi yangaphambili kwiindawo ezimbini zilandelayo:   "mutual information estimation", kunye   "neural networks" zokufunda "representation" phezu kwama "graphs". 1. 2. . Njengoko umgaqo we "InfoMax" [ ] ubongoza ukwandisa i "MI" phakathi kwe "inputs" kunye nee "outputs" ze "neural networks", iindlela ezininzi ezifana nee "ICA algorithms" [ ,  ] ziyazama ukusebenzisa ingcamango ye "MI" kwisifundo se "feature" esingakhokhelwanga. Nangona kunjalo, ezi ndlela azinakusetyenziswa kwi "deep neural networks" ngokulula ngenxa yobunzima bokubala i "MI" phakathi kwezinto ezininzi eziphilayo eziphilayo. Ngethamsanqa, i "Mutual Information Neural Estimation" (MINE) [ ] yenza ukulinganiswa kwe "MI" kwi "deep neural networks" kube nokwenzeka ngokuqeqesha inethiwekhi yobalo-manani njengomhlalutyi ukukhetha ii "samples" ezivela kwi "joint distribution" kunye ne "product of marginals" ezivela kuzo zombini izinto eziguquguqukayo. Ngokukodwa, i "MINE" isebenzisa ifomula ye "KL-based" ye "MI", ngelixa enye indlela engengawo ey "KL", i "Jensen-Shannon divergence" (JSD) [ ], ingasetyenziswa ngaphandle kokuxhalaba malunga nexabiso elichanekileyo le "MI". Ukulinganiswa kweMutual Information 3 1 21 2 30 . Ngophuhliso olukhawulezayo lwe "graph neural networks" (GNNs), inani elikhulu leendlela zokufunda ze "graph representation" ezisekelwe kwi "GNNs" ziphakanyiswe kule minyaka mihlanu idlulileyo, ezibonisa ukusebenza okunamandla kuneendlela zangaphambili ezisekelwe kwi "random walk" kunye ne "factorization" [ ,  ,  ,  ,  ]. Ngokuyintloko, ezi ndlela zingahlulwa zibe ngamacandelo axhaswe yikhohlileyo kunye nawo angaxhaswanga. Phakathi kwazo, kukho umsebenzi omningi malunga ne "supervised representation learning" phezu kwama "graphs" [ ,  ,  ,  ,  ]. Ngaphandle kokwahluka kwabo kwi "network architecture", bafumana impumelelo ngovavanyo ngoncedo lwe "labels" ezingafumaneki rhoqo kwiimeko zokwenyani. Kule meko, iindlela zokufunda ezingaxhaswanga yi "graph learning" [ ,  ,  ] zine "application potential" ebanzi. Indlela eyaziwayo yi "GraphSAGE" [ ], ifrem yokudibanisa ukuqeqesha ii "GNNs" ngomsebenzi osuselwe kwi "random-walk" kwisilungiselelo sayo esingaxhaswanga. Kwaye kutshanje, i "DGI" [ ] isebenzisa ingcamango ye "MI maximization" kwisithuba se "graph" kwaye ifumana ukusebenza okunamandla kwimodeli engaxhaswanga. Nangona kunjalo, i "DGI" isebenzisa ukwandisa okugubungayo ( ., ukwandisa i "MI" kwi "graph/patch-level") okwenza kube nzima ukugcina ulwazi olucokisekileyo kwigraph yangena. Ngaphandle koko, imeko efakwe kumsebenzi wokufunda osetyenziswa kwi "DGI" ibonakala ilucinezela kakhulu kwiimeko zokwenyani. Ngokuchaseneyo, sigxila ekususeni ukucinezela komsebenzi wokufunda kwaye sifikelele kwi "graphical mutual information maximization" kwi "node-level" ngokwandisa ngqo i "MI" phakathi kwe "inputs" kunye nee "outputs" ye "encoder". I "Representations" efunyenweyo yindlela yethu ingaqinisekanga ngakumbi ekugcineni ulwazi lwe "input graph", oluqinisekisa ubuchule bayo kwimisebenzi elandelayo ye "graph mining",  ., "node classification", "link prediction", kunye ne "recommendation". Izithuthi zobuchwepheshe kwi "graph representation learning" 6 15 33 34 39 7 9 25 40 48 11 16 41 16 41  Oko. umzekelo 3      I-GRAPHICAL MUTUAL INFORMATION: INKQASISO NOKUNGENZA KWAYO Ngaphambi kokuba siqhubeke, siqala sisongezele imiqondiso esetyenziswa kule nqaku. Kucha G = (V, E) ichaze igraph enama "nodes" angama "N" i "vi" ∈ V kunye nee "edges" i "eij" = ( "vi" , "vj" ) ∈ E. I "node features", kunye "empirical probability distribution" ethembekileyo, inikezelwa ngu "X" ∈ R  ×  = { "x" 1, · · · , "x" N } apho "xi" ∈ R  ichaza i "feature" yenqanawe "vi" . I "adjacency matrix" "A" ∈ R  ×  ibonisa ubudlelwane be "edge", apho "Aij" ehambelana nee "edge" "eij" ingaba yinani lokwenyani okanye i "multi-dimensional vector"1. N D D N N Injongo ye "graph representation learning" kukufunda i "encoder" "f" : R  ×  × R  ×  → R  × ′ , ukuze i "hidden vectors" "H" = { "h" 1, · · · , "h" N } = "f" ( "X" , "A" ) ibonise i "high-level representations" kubo bonke "nodes". Inkqubo yokucwangcisa ingabuyiselwa kwifom ye "node-wise". N D N N N D Ukubonisa oku, sichaza u "Xi" kunye no "Ai" kumntu "i" njengoko kuyiyo i "features" zabaswelana naye kunye ne "adjacency matrix" ehambelanayo ngokuxhomekeka kubaswelana nabo. Ngokukodwa, "Xi" ibandakanya zonke i "k-hop neighbors" ze "vi" kunye "k" ≤ "l" xa i "encoder" "f" iy "l"-layer GNN, kwaye ibandakanya inqanawe "i" ngokwayo ukuba siyongeze izixhumanisi zokuzenzekelayo kwi "adjacency matrix". Apha, sibiza i "sub-graph" eyenziwe ngu "Xi" kunye no "Ai" ngokuba yi "support graph" yenqanawe "i", ibhalwe njenge G  . Ngokuqulunqwa kwe "support graph", ukucwangciswa komntu ngamnye kuguquguquka ku "hi" = "f" (G  ) = "f" ( "Xi" , "Ai" ). i i  Kwi "Deep InfoMax" [ ], umsebenzi wokuqeqesha i "encoder" kukwandisa i "MI" phakathi kwe "inputs" kunye nee "outputs" zayo. I "MI" iyalinganiswa ngokusebenzisa inethiwekhi yobalo-manani njengomhlalutyi ukukhetha ii "samples" ezivela kwi "joint distribution" kunye nezo ziveliswe kwi "product of marginals". Ngokwemvelo, xa usebenzisa ingcamango ye "Deep InfoMax" kwii "graphs", kufuneka siyandise i "MI" phakathi kwe "representation" "hi" kunye ne "support graph" G  yenqanawe nganye. Sichaza olu hlobo lwe "graphical MI" njenge "I" ( "hi" ; G  ). Nangona kunjalo, akulula ukuchaza i I (hi ; Gi ). Iingxaki zezi: Iingxaki ekuchazeni i "graphical mutual information". 20 i i •   I "graphical MI" kufuneka ingachaphazeleki malunga nenombolo yenqanawe. Ngamanye amazwi, kufuneka sibe ne "I" ( "hi" ; G  ) = "I" ( "hi" ; G′  ), ukuba i G  kunye ne G′  ziyi "isomorphic" omnye komnye. i i i i •   Ukuba sisebenzisa i "MINE method" yokubala i "MI", umhlalutyi kwi "MINE" wamkela kuphela i "inputs" zosayizi olugqityiweyo. Oku akusebenzi kwii G  kuba ii G  ezahlukeneyo zihlala zibandakanya iinani ezahlukeneyo zabantu kwaye ngoko zinobungako obwahlukileyo. i i Ukujikeleza ingxaki yokuchaza i "graphical mutual information", eli candelo liqala ngokwazisa ingcamango ye "Feature Mutual Information" (FMI) ethembele kuphela kwi "node features". Emva kokuphefumlelwa kokuwohloka kwe "FMI", sibe sichaza i "Graphical Mutual Information" (GMI), ethatha zombini i "node features" kunye 3.1      I-FEATURE MUTUAL INFORMATION Sichaza i "empirical probability distribution" yee "node features" "Xi" njenge "p" ( "Xi" ), i "probability" ye "hi" njenge "p" ( "hi" ), kunye ne "joint distribution" nge "p" ( "hi" , "Xi" ). Ngokwe "information theory", i "MI" phakathi kwe "hi" kunye no "Xi" ichazwa njenge Ngokumangalisayo, sinalo olu luhlobo lwe "mutual information decomposition theorem" yokubala i "I" ( "hi" ; "Xi" ). Theorem 1 (Mutual Information Decomposition).      Ukuba i "conditional probability" "p" ( "hi" | "Xi" ) yi "multiplicative" (bona inkcazo ye "multiplicative" kwi [ ]), i "global mutual information" "I" ( "hi" ; "Xi" ) echazwe kwi Eq. 36 (1) ingahlulwa ibe yingqokelela enomlinganiselo we "local MIS", oko kukuthi,      ≤ "wij" ≤ 1  apho, "xj" yingcambu yes "j-th" yenqanawe "i", "in" lulwalamano lwayo yonke imicimbi kwi "Xi", kunye ne "weight" "wij" iyayithobela 1 kuwo wonke u "j". Ukubonisa le theorem, siqala sibonise ii "lemmas" ezimbini kunye nenkcazo enye. Lemma 1.  Kuzo naziphi na izinto eziguquguqukayo "X", "Y", kunye no "Z", sinako Ngokuwohluka kwi Theorem   singabala uhlanga olungasekhohlwayo lweEq.   nge "MINE" njengoko ii "inputs" zomhlalutyi ngoku ziba ziimbini ze ( "hi" , "xj" ) apho ubungako bugcina bungatshintshi ( ., "D" ′-by-"D"). Ngaphandle koko, singalungisa i "weights" ukubonisa indlela yokujikeleza kwee "input graphs". Umzekelo, ukuba "Xi" ibandakanya kuphela ii "one-hop neighbors" zenqanawe "i", ukubeka bonke i "weights" ilingana kuya kubangela i "MI" efanayo kwii "nodes" zangena ngeendlela ezahlukeneyo. 1, (2)  Oko. Ngaphandle kwezibonelelo zokuwohluka, kunzima ukuchaza amaxabiso ngokuchanekileyo kwee "weights" njengoko zinxulumene namaxabiso e "I" ( "hi" ; "xj" ) kunye nee "probability distributions" zazo. Indlela elula kukubeka bonke i "weights" babe ngu     , ngoko ke ukwandisa uhlanga olungasekhohlwayo lweEq.   kufana nokwandisa umda ongaphantsi we "I" ( "hi" ; "Xi" ), ngayo i "FMI" yokwenyani nayo yandiswa ngokuphathelele kwinqanaba elithile. 1 (2) Ukongeza kule ndlela, sinikezela ngekhambi eliphuculweyo ngokuthe ngqo ngokujonga ii "weights" njenge "trainable attentions", nto leyo engumbandela kwi "subsection" elandelayo.  Ephefumlelwe kukuwohluka kwi Theorem   sizama ukwakha i "trainable weights" kwelinye icala lama "graphs" ( ., 3.2      I-TOPOLOGY-AWARE MUTUAL INFORMATION 1,  Oko. indawo ye "topology") ukuze amaxabiso e "wij" angaba "flexible" kwaye abambe uphawu oluyimvelo lwama "graphs". Ekugqibeleni sichaza inkcazo ye "Graphical Mutual Information" (GMI). Inkcazo 2 (Graphical Mutual Information).   G = ( "X" , "A" )  I "MI" phakathi kwe "hidden vector" "h" kunye ne "support graph" yayo ichazwa njenge

Part of HackerNoon's growing list of open-source research papers, promoting free access to academic material.

Le audio iveliswe ngolwimi lokuqala lwebali!

Indlela Entsha Yokutshala U-AI Kudatha yeGrafu Ngaphandle Kokuqeqeshwa

About Author

IZIMVO

ZIJONGE IIMPAWU

ELI NQAKU LINIKEZELWE KU

Related Stories

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps