Strojno učenje na medicinskih podatkih z interpretabilnimi modeli

• Main Author: SROVIN, JURE VITO
• Other Authors: Kukar, Matjaž
• Format: Bachelor Thesis
• Language: Slovenian
• Published: 2021
• Subjects: interpretabilni modeli; RiskSLIM; strojno učenje v medicini; interpretable models; machine learning in medicine; sami
• Online Access: https://repozitorij.uni-lj.si/IzpisGradiva.php?id=125320; https://repozitorij.uni-lj.si/Dokument.php?id=140999&dn=; https://plus.si.cobiss.net/opac7/bib/54697731?lang=sl

Modeli strojnega učenja se uporabljajo v mnogih domenah, v katerih imajo napake lahko hude posledice na posameznika in družbo. Napačne klasifikacije in vzroke za njih je pogosto težko odkriti, še posebej pri uporabi kompleksnih modelov, katerih delovanje je človeku nerazumljivo. Cilj diplomske naloge je predstaviti pomen zmožnosti interpretacije modelov strojnega učenja in preveriti uspešnost delovanja preprostih modelov, ki so sami po sebi interpretabilni. Preizkusili smo RiskSLIM, algoritem za gradnjo redkih celoštevilskih linearnih modelov, ki so enostavni za uporabo, in ga primerjali z bolj priljubljenimi metodami strojnega učenja. Rezultate smo pridobili na dvorazrednih in večrazrednih medicinskih podatkovnih množicah različnih velikosti. Uspešnost modelov RiskSLIM na binarnih množicah je bila nekoliko slabša od preostalih metod, vendar še vedno zelo dobra. RiskSLIM ponuja odlično razmerje med interpretabilnostjo modela in uspešnostjo klasifikacij. Vendar pa deluje slabo na množicah, pri katerih je za uspešno klasifikacijo treba upoštevati veliko število atributov, kar pri RiskSLIM ni možno, saj je omejen na majhno število značilk. Uporabljamo ga lahko tudi na večrazrednih podatkovnih množicah s pomočjo metaklasifikatorjev. Njegova velika slabost je dolgotrajen postopek gradnje modela, ki se eksponentno podaljšuje z večjim številom atributov v množici. Zamudna je tudi ročna obdelava podatkov, saj je treba analizirati in diskretizirati vsako značilko posebej. Machine learning models are used in many domains where wrong decisions can have severe consequences for the individual and society. Misclassifications and their causes are often difficult to detect, especially when using complex models whose decision-making behaviour is unintelligible to humans. The goal of the thesis is to present the importance of interpretability of machine learning models and evaluate the performance of simple models that are inherently interpretable. We tested RiskSLIM, an algorithm for building sparse linear integer models that are easy to use and compared it to more popular machine learning methods. Results were obtained on binary and multiclass medical datasets of different sizes. The performance of RiskSLIM models on binary datasets was slightly worse than performance of other methods, but very good nonetheless. RiskSLIM offers an excellent trade-off between model interpretability and classification accuracy. However, it has poor performance on datasets, where a large number of features is required for successful classification, which is not possible with RiskSLIM, as it is limited to a small number of features per model. It can also be utilized on multiclass datasets using meta-classifiers. Its major drawback is the lengthy process of building the model, which is exponentially extended on datasets with large number of features. Manual data processing is also time consuming, as it is necessary to analyze and discretize each feature individually.