E-Locus - Institutional Repository of the University of Crete

Home Collections School/Department School of Sciences and Engineering Department of Computer Science Post-graduate theses

Post-graduate theses

Current Record: 58 of 824

[Add to Basket]

Identifier

000446856

Title

Automated feature engineering on relational data

Alternative Title

Αυτοματοποιημένη κατασκευή χαρακτηριστικών σε σχεσιακά δεδομένα

Author

Καλουρής, Δημήτριος Α.

Thesis advisor

Τσαμαρδινός, Ιωάννης

Reviewer

Χριστοφίδης, Βασίλης
Κομοντάκης, Νικόλαος

Abstract

Machine learning typically learns from a single table. However, in the age of big data it is often the case that data are distributed across many different tables in a relational database for efficiency. To work with relational data it is not rare for scientists perform feature engineering manually and intuitively. Additionally, many algorithms that produce a single table from a relational database have been proposed for this problem but none of them takes into account complex relational data schemas or they are limited in the paths they follow and the combinations of joins and aggregations they perform during feature generation. Moreover these algorithms, during feature generation, accumulate large number of features before performing feature selection and the feature selection algorithms are not optimized. To this end we created SRFGA a novel online feature engineering algorithm that performs joins and aggregations on the tables to create features and keeps only the most useful features, using the residuals calculated by a model to guide the feature selection. This algorithm can be used without any knowledge expertise, and it also unifies all the previous works in terms of visited paths and actions performed.

Language

English

Subject

Artificial intelligence

Feature construction

Relational databases

Σχεσιακές βάσεις

Τεχνητή νοημοσύνη

Issue date

2022-03-18

Collection

School/Department--School of Sciences and Engineering--Department of Computer Science--Post-graduate theses