Javascript must be enabled for the correct page display

Preprocessing Pipeline Meta-Learning Using Reliable and Diverse Meta-Target Selection

Albietz, Charlie (2023) Preprocessing Pipeline Meta-Learning Using Reliable and Diverse Meta-Target Selection. Master's Thesis / Essay, Artificial Intelligence.


Download (1MB) | Preview
[img] Text
Restricted to Registered users only

Download (124kB)


In this paper, a meta-learning system was trained to predict the performance of dataset preprocessing (DPP) pipelines, allowing for automation of the DPP pipeline optimisation process. Previous meta-learning studies have highlighted a need for improvements concerning meta-model accuracy when predicting ideal DPP pipelines. For this reason, the current project aims to employ a new method that is expected to help increase meta-model accuracy for DPP recommendation. The method that is proposed aims to improve how meta-targets are selected. Usually, meta-targets used for DPP pipeline recommendation consist of pipelines with high average performances across all previously evaluated tasks: reliable pipelines. An alternative approach for meta-target selection is proposed: the principle of diversity. Here meta-targets are chosen that maximise performance for subgroups of tasks. The current meta-learner was trained and tested on classification and regression datasets using reliable and diverse meta-target selection methods. The results show that the meta-learning accuracy increased when using diverse meta-targets instead of reliable meta-targets for classification datasets. Furthermore, this model was compared to a popular DPP pipeline optimisation tool, TPOT. Additionally, the results suggest that the reliable meta-target selection method leads to a better DPP pipelines than the diverse meta-target selection method.

Item Type: Thesis (Master's Thesis / Essay)
Supervisor name: Jaeger, H.
Degree programme: Artificial Intelligence
Thesis type: Master's Thesis / Essay
Language: English
Date Deposited: 06 Apr 2023 12:19
Last Modified: 06 Apr 2023 12:19

Actions (login required)

View Item View Item