Ekstremno naključni kvantilni gozdovi

BALON, MATJAŽ (author), Kononenko, Igor (mentor)

Ekstremno naključni kvantilni gozdovi so ansambelska metoda, ki z vnašanjem dodatne naključnosti in kvantilov razširja običajne naključne gozdove. V tem delu smo preverili njeno napovedno točnost z različnimi merami uspešnosti ter hitrost izvajanja. Analizirali smo tudi vplive različnih parametrov in velikosti učne množice na uspešnost delovanja in čas potreben pri izvajanju. Metodo smo primerjali z različnimi merami uspešnosti in časi delovanja še z dvema metodama, ki za napovedovanje uporabljata kvantile in s tremi metodami, ki dajejo napovedne intervale. Ekstremno naključni kvantilni gozdovi so se izkazali kot zelo konkurenčni tako v točnosti napovedovanja po različnih merah kot tudi v hitrosti izvajanja.

• Language: Slovenian
• Keywords: strojno ucenje, kvantili, nakljucni gozdovi, regresija
• Work type: Undergraduate thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2016
• PID: 20.500.12556/RUL-82724
• Publication date in RUL: 19.05.2016

Secondary language

• Language: English
• Title: Extremely randomized quantile forests

Abstract

Extremely randomized quantile forests are an ensemble method which extends ordinary random forests with additional randomness and quantiles. In this work we checked its prediction accuracy with different success measures and execution speed. We also analyzed influences of various parameters and size of learning dataset on prediction performance and time needed for execution. We compared method with different success measures and execution time also with two methods that use quantiles for prediction and three methods that give prediction intervals. Extremely randomized quantile forests have been proved as competitive in terms of prediction strength with different measures and also in speed of execution.

• Keywords: machine learning, quantiles, random forest, regression