A Unified Model Representation of Machine Learning Knowledge

A Uni ied Model Rep esen a ion o Machine Lea ning Knowledge J. G. En ´ ıquez1,⇤, A. Ma ´ ınez-Rojas1, D. Lizcano2 and A. Jim´ enez-Ram´ ı ez1 1Compu e Languages and Sys ems Depa men . Escuela T´ ecnica Supe io de Ingenie ´ ıa In o m´ a ica, A enida Reina Me cedes, s/n, 41012, Se illa. Spain 2Uni e sidad a dis ancia de Mad id. Ca e e a de La Co u˜ na, KM.38,500, ´ ıa de Se icio, no 15, 28400, Collado Villalba, Mad id. Spain E-mail: [email p o ec ed] ⇤Co esponding Au ho Recei ed 20 Decembe 2019; Accep ed 14 Ap il 2020; Publica ion 03 June 2020 Abs ac Nowadays, Machine Lea ning (ML) algo i hms a e being widely applied in i ually all possible scena ios. Howe e , de eloping a ML p ojec en ails he e o o many ML expe s who ha e o selec and con igu e he app op ia e algo i hm o p ocess he da a o lea n om, be ween o he hings. Since he e exis housands o algo i hms, i becomes a ime-consuming and challenging ask. To his end, ecen ly, Au oML eme ged o p o ide mechanisms o au oma e pa s o his p ocess. Howe e , mos o he e o s ocus on applying b u e o ce p ocedu es o y di e en algo i hms o con igu a ion and selec he one which gi es be e esul s. To make a sma e and mo e e icien selec ion, a eposi o y o knowledge is necessa y. To his end, his pape p oposes (1) an app oach owa ds a common language o consolida e he cu en dis ibu ed knowledge sou ces ela ed he algo i hm selec ion in ML, and (2) a me hod o join he knowledge ga he ed h ough his language in a uni ied s o e ha can be exploi ed la e on, and (3) a aceabili y links main enance. The p elimina y e alua ions o his app oach allow o c ea e a Jou nal o Web Enginee ing, Vol. 19 2, 319–340. doi: 10.13052/jwe1540-9589.1929 © 2020 Ri e Publishe s 320 J. G. En ´ ıquez e al. uni ied s o e collec ing he knowledge o 13 di e en sou ces and o iden i y a bunch o esea ch lines o conduc . Keywo ds: Machine Lea ning, Au oma ed Machine Lea ning, Knowledge Rep esen a ion, Model-D i en Enginee ing. 1 In oduc ion Machine Lea ning (ML) en ails he s udy o algo i hms ha au oma ically imp o e h ough expe ience [18]. This kind o algo i hms has been suc- cess ully and b oadly applied in he pas [19] and nowadays is ecei ing inc easing a en ion due o he a o dable access o bigge compu a ion powe o machines. A ML p ojec equi es selec ing an app op ia e algo i hm o p ocess he da a o lea n om, which is ypically named c ea ing he da a model. Howe e , he e a e housands o algo i hms unde he pa adigm o ML, each o hem ailo ed o some speci ic asks o con ex s. In addi ion, many o hese algo i hms o e a di e en se o pa ame e s o be con igu ed (e.g., selec ing he numbe o laye s in a neu al ne wo k). Many exis ing app oaches ocus on he la e ask, i.e., suppo ing he use a e he algo i hm selec ion is done, and ew o hem ecommend an algo i hm always a e he use has p o ided he da ase . As an example, he ecen esea ch a ea o Au oML [28] aims o au oma e he di e en s eps o ML p ojec s. None heless, such app oaches neglec he ea ly s ages o he p ojec . Many o hem jus p o ide a b u e o ce mechanism ha uns se e al algo i hms in la e s ages o he p ojec , i.e., when he da ase is eady. Thus, li le e o has been done o suppo he use in he algo i hm selec ion in an e icien manne (i.e., wi hou applying b u e o ce) and based on he p oblem cha ac e is ics (i.e., he ea ly in o ma ion). The algo i hm selec ion is speci ically challenging since he exis ing knowledge ega ding his ask is dis ibu ed ac oss di e en sou ces and each o hem is speci ied in a non-s anda d manne , hus, making i di icul o con- solida e in o ma ion om di e en sou ces, i.e., he name o he algo i hms —o amily o algo i hms—, he selec ion c i e ia, and he cha ac e is ics o he p oblem ha a ec he selec ion a e he e ogeneous (c . Figu e 1). To educe he isk o aking inaccu a e decisions due o a lack o in o ma- ion, a cen al eposi o y o he ML Knowledge which s o es he in o ma ion in a s uc u ed way is equi ed. In o de o add ess his p oblem, his pape p oposes (c . Figu e 2), on he one hand, a uni ied language o ep esen ing