On the Stability of Feature Selection in the Presence of Feature Correlations

Sechidis, Konstantinos; Papangelou, Konstantinos; Nogueira, Sarah; Weatherall, James Owen; Brown, Gavin

doi:10.1007/978-3-030-46150-8_20

Cited by 14 publications

(14 citation statements)

References 11 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The Kuncheva index (KI) and the φ measure plainly ignore possible correlations between features. Inspired from Sechidis [9], we consider a scenario where a selection algorithm toggles, between pairs of correlated variables, across runs. More specifically, let us represent the selection matrix Z below with one row per selection run, z i,f indicating the selection of feature f in run i…”

Section: Related Workmentioning

confidence: 99%

“…Sechidis [9] generalizes φ in order to accurately measure the selection stability in the presence of highly correlated variables:…”

Section: Related Workmentioning

confidence: 99%

“…This shared importance is multiplied by the similarity s f,g between the two features. The feature f can share its importance with several features of run j, but its total shared importance cannot exceed its own importance I i,f according to the constraint (9) (and its reciprocal (10)). Feature importance values are normalized such that f ∈Fi I i,f = k, the average number of selected features.…”

Section: Stability As Maximal Shared Importancementioning

confidence: 99%

“…Condition 1 is satisfied as the coefficients multiplying x f,g in the set of constraints ( 9) and ( 10) are 0 or 1. Condition 2 holds because each variable x f,g has a non-zero coefficient in two constraints, one from the set of constraints (9) and one from (10). Condition 3 holds as we let B be the rows of A representing the set of constraints (9) and C be the rows of A corresponding to the set of constraints (10).…”

Section: Stability As Maximal Shared Importancementioning

confidence: 99%

“…This could lead to a pessimistic estimation of the stability when some features are selected in one run and other features, distinct from but highly correlated to the initial ones, are selected in another run. Previous works specifically address this issue by proposing stability indices taking into account feature correlations (or, more generally, feature similarity values) [9,12].…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Robust Selection Stability Estimation in Correlated Spaces

Hamer

Dupont

2021

Machine Learning and Knowledge Discovery in Databases. Research Track

View full text Add to dashboard Cite

The stability of feature selection refers to the variability of the selected feature sets induced by small changes of data sampling or analysis pipeline. Instability may strongly limit a sound interpretation of the selected features by domain experts. This work addresses the problem of assessing stability in the presence of correlated features. Correctly measuring selection stability in this context amounts to estimate to which extent several correlated variables contribute to predictive models, and how such contributions may change with the data sampling. We propose here a novel stability index taking into account such multivariate contributions. The shared contributions of several variables to predictive models do not only depend on the possible correlations between them. Computing this stability index therefore requires to solve a weighted bipartite matching problem to discover which variables actually share such contributions. We demonstrate that our novel approach provides more robust stability estimates than current measures, including existing ones taking into account feature correlations. The benefits of the proposed approach are demonstrated on simulated and real data, including microarray and mass spectrometry datasets. The code and datasets used in this paper are publicly available: https://github.com/hamerv/ecml21.

show abstract

Section: Related Workmentioning

confidence: 99%

“…Sechidis [9] generalizes φ in order to accurately measure the selection stability in the presence of highly correlated variables:…”

Section: Related Workmentioning

confidence: 99%

Section: Stability As Maximal Shared Importancementioning

confidence: 99%

Section: Stability As Maximal Shared Importancementioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%