2025-02-192025-02-192024-06-07Jiménez Navarro, M.J., Martínez Ballesteros, M.d.M., Brito, I.S., MartÍnez-Álvarez, F. y Asencio-Cortés, G. (2024). Embedded feature selection for neural networks via learnable drop layer. Logic Journal of the IGPL. https://doi.org/10.1093/jigpal/jzae062.1367-07511368-9894https://hdl.handle.net/11441/168969Featureselectionisawidelystudiedtechniquewhosegoalistoreducethedimensionalityoftheproblembyremovingirrelevantfeatures.Ithasmultiplebenefits,suchasimprovedefficacy,efficiencyandinterpretabilityofalmostanytypeofmachinelearningmodel.Featureselectiontechniquesmaybedividedintothreemaincategories,dependingontheprocessusedtoremovethefeaturesknownasFilter,WrapperandEmbedded.Embeddedmethodsareusuallythepreferredfeatureselectionmethodthatefficientlyobtainsaselectionofthemostrelevantfeaturesofthemodel.However,notallmodelssupportanembeddedfeatureselectionthatforcestheuseofadifferentmethod,reducingtheefficiencyandreliabilityoftheselection.Neuralnetworksareanexampleofamodelthatdoesnotsupportembeddedfeatureselection.Asneuralnetworkshaveshowntoprovideremarkableresultsinmultiplescenariossuchasclassificationandregression,sometimesinanensemblewithamodelthatincludesanembeddedfeatureselection,weattempttoembedafeatureselectionprocesswithageneral-purposemethodology.Inthiswork,weproposeanovelgeneral-purposelayerforneuralnetworksthatremovestheinfluenceofirrelevantfeatures.TheFeature-AwareDropLayerisincludedatthetopoftheneuralnetworkandtrainedduringthebackpropagationprocesswithoutanyadditionalparameters.Ourmethodologyistestedwith17datasetsforclassificationandregressiontasks,includingdatafromdifferentfieldssuchasHealth,EconomicandEnvironment,amongothers.Theresultsshowremarkableimprovementscomparedtothreedifferentfeatureselectionapproaches,withreliable,efficientandeffectiveresults.application/pdf25 P.engAttribution-NonCommercial-NoDerivatives 4.0 Internationalhttp://creativecommons.org/licenses/by-nc-nd/4.0/Feature selectionNeural networkClassificationRegressionEmbedded feature selection for neural networks via learnable drop layerinfo:eu-repo/semantics/articleinfo:eu-repo/semantics/openAccess10.1093/jigpal/jzae062