Please use this identifier to cite or link to this item: https://rfos.fon.bg.ac.rs/handle/123456789/1650
Full metadata record
DC FieldValueLanguage
dc.creatorMišulić, Danijel
dc.creatorJovanović, Miloš
dc.date.accessioned2023-05-12T11:07:08Z-
dc.date.available2023-05-12T11:07:08Z-
dc.date.issued2017
dc.identifier.issn1451-4397
dc.identifier.urihttps://rfos.fon.bg.ac.rs/handle/123456789/1650-
dc.description.abstractProblematika predviđanja učinka košarkaša na utakmicama putem algoritama mašinskog učenja sve više dobija na interesovanju i značaju u poslednje vreme. Pregled osnovne ideje rada da je na konkretnom primeru predviđanja poena NBA košarkaša koristeći vremenske serije, a zatim modelovanje njihovog učinka na osnovu predviđenog učinka njihovih saigrača. Pristup koji je korišćen za više-izlaznu (multi-target) predikciju je steking (eng. stacking). Cilj rada je pokazati kako steking pristup može uticati na poboljšanje performansi prediktivnih modela. U ovom radu su kreirana i dva modela: nezavisni model koji će na osnovu stvarnih vrednosti predviđjati određeni izlaz i drugi, više-ciljni (eng. multi-target) model, koji će korisititi predviđene vrednosti osnovnog modela i vršiti korekciju tih vrednosti. Prvobitni model predviđa vrednosti na nivou pojedinačnih opservacija dok drugi model dalje koristi te predviđene vrednosti korišćenjem multi-target pristupa kako bi identifikovao međuzavisnosti između izlaznih vrednosti prvobitnog modela. Dati modeli se ukrštaju na taj način što će izlaz nezavisnog modela biti korišćen kao ulaz u multi-target model. Eksperimentalna evaluacija je pokazala da multi-target model daje bolje rezultate nego osnovni na osnovu poređenja srednje apsolutne greške osnovnog i datog modela. Dobijeni preliminarni rezultati ukazuju na korisnost ovakvog pristupa i motivišu širu zajednicu u daljem razvoju sličnih metoda koje se zasnivaju na multi-target regresiji i stacking pristupu.sr
dc.description.abstractThe problem of anticipating the impact of basketball players in games through machine learning algorithms is gaining more and more interest in recent times. An overview of the basic idea of the work will be given on the concrete case of predicting the points of the NBA basketball players using time series, and then modeling their performance based on the predicted performance of their teammates. The approach used for multi-target prediction is stacking. The aim of the paper is to show how a stacking approach can improve performance of prediction models. Two models are created and evaluated: an independent model that, based on real values, predicts a certain output, and another, multi-target model, which will use the predicted values of the basic model and make some correction of these values. The original model predicts values at the level of individual observations while the other model uses these predicted values using a multi-target approach to identify interdependencies between the output values of the original model. The given models are crossed in such way that the output of the independent model will be used as an input to the multi-target model. Experimental evaluation showed that the multi-target model yields better results than the basic one based on the comparison of the mean absolute error of the basic and given model. The obtained preliminary results indicate the usefulness of this approach and motivate the wider community in the further development of similar methods based on multi-target regression and stacking approach.en
dc.publisherUniverzitet u Beogradu - Fakultet organizacionih nauka, Beograd
dc.rightsopenAccess
dc.sourceInfo M
dc.subjectvremenske serijesr
dc.subjectsportsr
dc.subjectpredviđanjesr
dc.subjectmulti-target regresijasr
dc.subjectmašinsko učenjesr
dc.subjectkošarkasr
dc.subjecttime seriesen
dc.subjectsporten
dc.subjectpredictionen
dc.subjectmulti-target regressionen
dc.subjectmachine learningen
dc.subjectbasketballen
dc.titleViše-izlazna regresija za predviđanje učinka košarkašasr
dc.titleMulti-target regression for predicting basketball player performanceen
dc.typearticle
dc.rights.licenseARR
dc.citation.epage7
dc.citation.issue63
dc.citation.other16(63): 4-7
dc.citation.rankM52
dc.citation.spage4
dc.citation.volume16
dc.identifier.rcubconv_746
dc.type.versionpublishedVersion
item.cerifentitytypePublications-
item.fulltextNo Fulltext-
item.grantfulltextnone-
item.openairetypearticle-
item.openairecristypehttp://purl.org/coar/resource_type/c_18cf-
Appears in Collections:Radovi istraživača / Researchers’ publications
Show simple item record

Google ScholarTM

Check


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.