Chii chaitika
Vatsvagiri Mulin Tian naAjitesh Srivastava vakaunza G2I, nhanho mbiri-nhanho yekugadzira kupindira hypotheses kubva kumagirafu neural network. Iro bepa rinodzokorora tsananguro sedambudziko rekupindira-dhizaini: kutanga kuwana shanduko dzenzvimbo dzinogona kushandura kufanotaura, wozosarudza yakarongwa seti yemitemo padanho retiweki.
Bepa racho, rakatumirwa kuarXiv muna Aug. 24, rinopa G2I senzira yekutsanangura fungidziro dzakaitwa negraph neural network uye kushandura tsananguro idzodzo mukupindira kunobvira. Graph neural network yakagadzirirwa data rehukama, uko ruzivo rwakakodzera runosanganisira kubatana pakati pevanhu, masangano kana mamwe masangano. Vanyori vanopokana kuti nzira dzakawanda dziripo dzekutsanangura dzinoshanda zvakanyanya padanho remunhu-node uye hazvitsigire zvakaringana kurongwa kwekupindira kwetiweki. Iyo arXiv rekodhi zvakare inonyora bepa sekugamuchirwa ku35th ACM International Musangano weRuzivo uye Ruzivo Management muna 2026.
G2I ine matanho maviri. Padanho renzvimbo, kutsvaga kwemakaro kunoratidza izvo vanyori vanotsanangura sezvishoma, zvinogoneka shanduko kune node maficha uye nevavakidzani-nhanho mamiriro ayo anogona kuburitsa mhedzisiro. Iwo ekunyepedzera anozoshandurwa kuita mitemo inotarirwa kunzwisiswa kune nyanzvi yedomeni inogona kunge isiri nyanzvi yeAI. Iyo nzira inounzwa seimwe nzira yekuenderera mberi nekugadziriswa kwemask, nzira inoshandiswa nevatsananguri venhema vanosanganisira CF-GNNExplainer uye CF². Vanyori vanoti nzira dzakavakirwa nemasiki dzinogona kufungidzira kuti mipendero inobatika, kushandisa simba pane isingachinjiki kana neimwe nzira isingaite hunhu, uye inoda yakakura computation.
Padanho retiweki, G2I inobata kupindira kusarudzwa sedambudziko-rakaganhurirwa rekuvharidzira rinoratidzwa mune disjunctive yakajairika fomu, kana DNF. Pepa rinotsanangura chinangwa chekuvhara sechisingadzike uye chinenge chiri pasi modular, zvivakwa zvinobvumira maitiro emakaro ekusarudza ane vimbiso dzedzidziso. Vanyori vanorondedzera zviedzo pamagirafu ekugadzira uye network-yepasirese yekuzviuraya-njodzi uye vanoti G2I yakagadzira scalable, inodhura-inoshanda nzira dzekupindira nehunyanzvi hwakanyanya kuvandudzwa pamusoro pemasiki-based counterfactual nzira. Iyo sosi haipe iwo abstract's manhamba mhedzisiro, saizi kana hunhu hwe dataset, iyo chaiyo yekupindira mitemo, kana computational marongero anoshandiswa kuenzanisa.
Zvakatorwa pamwechete, nhanho mbiri idzi dzinobatanidza kufunga kwenzvimbo yekunyepedzera netiweki-level sarudzo. Zano remunharaunda rinotsanangura shanduko inogona kutenderedza node, nepo nhanho yekuvhara inotevera ichifunga kuti seti yemitemo inogona kusarudzwa sei pasi pebhajeti shoma. Bepa rinopa chimiro ichi senzira yekuita kuti kufungidzira kupindira kuve kuite uye kuwedzereka. Kuenzanisa kwayo kunoramba kwakatarisana nekubudirira kunopesana neakadudzwa mask-based maitiro, nepo sosi ichisiya nhamba yakawanikwa uye ruzivo rwekuita zvisina kutaurwa. Iyi miganho inotsanangura zvinogona kupedzwa kubva pane zvakashumwa zviedzo.
Nei zvichikosha
Graph-based AI iri kuwedzera kushandiswa kudzidza vanhu vakabatana uye masisitimu, asi tsananguro yekufanotaura haisi yega kupindira kunoshanda. G2I inoitirwa kuburitsa mitemo yakareruka, inobatika uku ichidzikisa mutoro we computational uye shanduko dzisingagoneke dzakabatana nevamwe varipo vatsananguri venhema.
Dambudziko rinoshanda rakagadziriswa neG2I ndiro mukaha uri pakati pekufanotaura njodzi uye kusarudza kuti ndechipi chiito chingashandura njodzi iyoyo. Mutevedzeri unogona kuona node seyakakosha kana kupa mukana wakafanotaurwa mukuru pasina kuudza nyanzvi yezveutano kana magariro esainzi kuti shanduko dzinogoneka, ndedzipi vatambi dzinogona kuita, kana kuti kupindira kwakati kunofanira kukosheswa sei pasi pezviwanikwa zvishoma. Nekugadzira tsananguro sedhizaini yekupindira, bepa rinotarisa kune iyo mibvunzo yekushandisa pane kubata kududzira semagumo pachayo.
Iko kusimbiswa kweunhu hunogona kuita kwakakosha nekuti data yegirafu inogona kuva nemusanganiswa wemamiriro ezvinhu anoshanduka, hunhu hwakagadziriswa uye hukama hungave hwakaoma kana husina kunaka kuchinja. Sosi yacho inoshoropodza nzira dzinogona kugovera nhamburiko kune isingachinjiki kana isingaite hunhu. Mitemo yakarongwa yeG2I saka inogona kupa hwaro hwakajeka hwekuongororwa kwenyanzvi, kunyanya kana muiti wesarudzo achifanira kuenzanisa nhamba shoma yekuchinja kunogoneka pane kuongorora mask yekugadzirisa yakaoma. Yayo network-level yekuvhara gadziriso inogadzirisawo chokwadi chinoshanda chekuti chirongwa chekupindira kazhinji chine bhajeti uye chinofanirwa kusvika kune anopfuura munhu mumwechete kana node.
Iko kukosha kweruzhinji kunofanirwa kuwedzeredzwa. Bepa rinoshuma nzira yekugadzira fungidziro dzekupindira, kwete humbowo hwekuti kupindira kwacho kunokonzera mhedzisiro iri nani. Tsananguro yekupokana kubva kune inofanotaura modhi inogona kutsanangura izvo zvingashandura kuburitsa kweiyo modhi pasina kumisikidza hukama hunokonzera muhuwandu hwevanhu. Musiyano iwoyo unonyanya kukosha munzvimbo yenjodzi yekuzviuraya yakadanwa nekwakabva, uko manyepo, manyepo, zvinonetsa kuvanzika uye kusaenzana kurapwa kunogona kuve nemhedzisiro. Kwakabva hakutauri kuti G2I yakaendeswa, yakaongororwa nenharaunda dzakakanganisika, kana kuratidzwa kuvandudza sarudzo dzekiriniki kana yeruzhinji-hutano.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
Which component of an AI application is the machine-learning model itself?
Zvekutarisa zvinotevera
Mhedzisiro yebepa yakashumwa inongogumira pakuedza pamagirafu ekugadzira uye chaiyo-yepasirese yekuzviuraya-njodzi network, uye iyo abstract inopa hapana manhamba ekuita manhamba. Kumwe kuongorora kunofanirwa kuyedza kuti nyanzvi dzemadomasi dzinowana here mitemo inonzwisisika, kana shanduko dzakarongwa dzichigoneka uye dzakaringana, uye kana dzichivandudza mhedzisiro kana dzichishandiswa mukuita.
Muedzo unotevera ndewekuti mitemo yeG2I inoramba ichibatsira kunze kwekuedzwa kwepepa. Iyo sosi inozivisa magirafu ekugadzira uye network-yepasirese-njodzi yekuzviuraya, asi haitsananguri maitiro ekuunganidza data, huwandu hwevanhu vanomiririrwa, tsananguro dzengozi, zviripo kana kuti hukama hwetiweki hwakavakwa sei. Iwo madonhwe ndiwo achaona kana iyo yakashumwa kubudirira kunoratidza nzira inobatsira yakafararira kana mhedzisiro inoenderana nemamwe magirafu zvimiro uye fungidziro.
Kuongorora kwakazvimiririra kunofanirwa kuongorora kunaka uye kugona kwezvirongwa zvakarongwa, kwete chete kana vachichinja fungidziro yemuenzaniso. Nyanzvi dzeDomain dzinogona kuongorora kana mitemo ichinzwisisika, ndeyechokwadi uye inopindirana nemaitiro anogamuchirwa. Vatsvaguri vanofanirwa kuyedzawo kusimba kwekushaikwa kana ruzha zvinongedzo, shanduko mubhajeti rekupindira, mamwe magirafu zvimiro uye mhando dzakasiyana. Tsime rinoti kutsvaga kwemakaro kune vimbiso mune mamwe mamiriro uye kuti iwo mamiriro akange akakwana empirically, asi haitsanangure mamiriro ezvinhu kana kutsanangura kuti mhedzisiro yacho inonzwa sei kana vakundikana.
Zvinokosha zvisingazivikanwe zvinosanganisira kunyatsoita budiriro pamusoro peCF-GNNExplainer neCF², mutengo wekuverengera wenzira yega yega, huwandu hwemagirafu nemakesi akaedzwa, uye kuti G2I inoshandura chokwadi chemodhi kana nzira yekutsanangura chete. Iyo abstract zvakare haitaure kana kodhi kana dhataseti iripo, ingave chidzidzo ichi chinosanganisira kurongeka kana kuongororwa kwekuvanzika, kana kuti chero mutemo wakatsanangurwa wakaedzwa sekupindira chaiko. Kugamuchirwa kwebepa kuCIKM chiitiko chekushambadzira, kwete chisimbiso chakazvimirira chekuti nzira yacho ndeyechokwadi kana kuti yakagadzirira kushandiswa muzvisarudzo zvepamusoro.