EURO 2024 Copenhagen
Abstract Submission

EURO-Online login

2388. Feature set aggregation using community detection in networks

Invited abstract in session MD-31: Network Analytics, stream Analytics.

Monday, 14:30-16:00
Room: 046 (building: 208)

Authors (first author is the speaker)

1. Christoph Lohrmann
Department of Computer Science, Reykjavik University
2. María Óskarsdóttir
Department of Computer Science, Reykjavik University

Abstract

In supervised machine learning, especially with high dimensional data, feature selection is an important approach that aims to ensure efficient computation, to achieve a high generalization ability of the learning algorithm, and to support the interpretability of the results. However, the feature subsets selected during feature selection may vary depending on the partition of the training dataset used or the feature selection method deployed. This instability can be a result of including redundant and irrelevant features in some of these subsets or not retaining all relevant features in each of them. The goal of this study is to develop a methodology to find a subset of the relevant features using network science and community detection from among the individual feature subsets. The individual feature subsets will be used to construct a network, where each node represents a feature, and edges between nodes reflect whether two features were selected into the same subset. By considering several runs of a feature selection method, features that appear together often will have a stronger connection, represented by a higher weight on the edge between them. This network will then be analyzed to discover communities, which ideally represent a set of features that is highly connected and, thus, is jointly highly important for the classification task. Optimally, this also allows to visually and/or numerically distinguish the sets of relevant, redundant, and irrelevant features.

Keywords

Status: accepted

Back to the list of papers