Analiza različitih metoda predviđanja vremena do događaja

master's thesis

• Title (english): Analysis of various methods for predicting time to event
• Author: Petra Dunja Grujić Ostojić
• Mentor: Damir Pintar (mentor)
• Committee member: Damir Pintar (predsjednik povjerenstva)
• Committee member: Mihaela Vranić (član povjerenstva)
• Committee member: Luka Humski (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2024-07-01, Croatia

Abstract

Zadatak predviđanja vremena do događaja razlikuje se od uobičajenih problema regresije i klasifikacije gdje je cilj predviđanje kontinuirane ili diskretne varijable. Osnovni razlog nemogućnosti primjene tradicionalnih metoda regresije i strojnog učenja jest prisutnost cenzuriranih podataka.

U prvom dijelu rada predstavljeni su osnovni koncepti analize doživljenja i dan je pregled tradicionalnih metoda, ali i prilagođenih modela strojnog učenja. U drugom dijelu rada su primijenjene različite metode analize doživljenja na dva javno dostupna podatkovna skupa. U German Breast Cancer Study predviđa se vrijeme do povratka raka dojke, a u podatkovnom skupu Bank Customer Churn vrijeme do napuštanja banke. Od tradicionalnih statističkih modela razvijen je Coxov regresijski model proporcionalnih hazarda. Od ansambla modela u radu su korišteni: model slučajnih šuma, model ekstremno slučajnih šuma i 2 modela temeljena na gradijentnom poboljšanju. Također, isprobani su i modeli temeljeni na stroju potpornih vektora. Na manjem podatkovnom skupu koji opisuje povratak raka dojke najboljim modelima su se pokazali statistički razvijen Coxov model i model ekstremno slučajnih šuma. Na većem podatkovnom skupu koji opisuje napuštanje banke najboljim modelom se pokazao model slučajnih šuma. Općenito, moglo se primijetiti da ansambli modela ostvaraju bolje rezultate na većem podatkovnom skupu. Prethodno usporedbi metoda predviđanja vremena do događaja, napravljena je eksploratorna analiza podatkovnih skupova koja daje uvid u značajke koje povećavaju odnosno smanjuju hazard.

Abstract (english)

The task of time-to-event prediction differs from traditional regression and classification problems, which aim to predict a continuous or discrete variable. The main reason traditional regression and machine learning methods are inappropriate for this task is the presence of censored data.

The first part of the thesis presents the basic concepts of survival analysis and provides an overview of traditional methods, as well as adapted machine learning models. In the second part of the thesis, various survival analysis methods were applied to two publicly available datasets. In the German Breast Cancer Study, the aim is to predict the time to breast cancer recurrence, and in the Bank Customer Churn dataset, the time to bank client churn. Among traditional statistical models, the Cox proportional hazards regression model was developed. Among ensemble models, the following were used: Random Survival Forest, Extra Survival Trees, and two models based on Gradient Boosting. Additionally, models based on Support Vector Machines were also tested. For the smaller dataset describing breast cancer recurrence, the best models were the statistically developed Cox proportional hazards model and the Extra Survival Trees. For the larger dataset describing bank churn, the best model was the Random Survival Forest. Generally, it could be observed that ensemble models achieved better results on the larger dataset. Prior to comparing time-to-event prediction methods, an exploratory data analysis was conducted to provide insight into the features that increase or decrease the hazard.

• Keywords: analiza doživljenja; analiza vremena do događaja; predviđanje vremena do događaja; cenzura; hazard; funkcija doživljenja; Coxov model proporcionalnih hazarda; model slučajnih šuma; model ekstremno slučajnih šuma; modeli predviđanja vremena do događaja temeljeni na gradijentnom poboljšanju; modeli predviđanja vremena do događaja temeljeni na stroju potpornih vektora; predviđanje vremena do povratka raka dojke; predviđanje vremena do napuštanja banke
• Keywords (english): survival analysis; time-to-event analysis; time-to-event prediction; censorship; hazard; survival function; Cox proportional hazards model; Random Survival Forest; Extra Survival Trees; Gradient Boosted Time-To-Event Models; Survival Support Vector Machine Models; time to breast cancer recurrence prediction; time to bank client churn prediction
• Language: croatian
• URN:NBN: urn:nbn:hr:168:723726
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2024-10-12 20:44:45