Being able to compare machine learning models in terms of performance is a fundamental part of improving the state of the art in a field. However, there is a risk of getting locked into only using a few -- possibly not ideal -- performance metrics, only for comparability with earlier works. In this work, we explore the possibility of reconstructing new classification metrics starting from what little information may be available in existing works. We propose three approaches to reconstruct confusion matrices and, as a consequence, other classification metrics. We empirically verify the quality of the reconstructions, drawing conclusions on the usefulness that various classification metrics have for the reconstruction task.

RECLAIM: Reverse Engineering Classification Metrics / Giobergia, Flavio; Baralis, Elena. - (2022), pp. 106-113. ((Intervento presentato al convegno IEEE International Conference on AI and Knowledge Engineering tenutosi a 19-21 September 2022 nel Laguna Hills, CA (USA) [10.1109/AIKE55402.2022.00024].

RECLAIM: Reverse Engineering Classification Metrics

Giobergia, Flavio;Baralis, Elena
2022

Abstract

Being able to compare machine learning models in terms of performance is a fundamental part of improving the state of the art in a field. However, there is a risk of getting locked into only using a few -- possibly not ideal -- performance metrics, only for comparability with earlier works. In this work, we explore the possibility of reconstructing new classification metrics starting from what little information may be available in existing works. We propose three approaches to reconstruct confusion matrices and, as a consequence, other classification metrics. We empirically verify the quality of the reconstructions, drawing conclusions on the usefulness that various classification metrics have for the reconstruction task.
File in questo prodotto:
File Dimensione Formato  
RECLAIM__IEEE_AIKE_-3.pdf

accesso aperto

Tipologia: 2. Post-print / Author's Accepted Manuscript
Licenza: PUBBLICO - Tutti i diritti riservati
Dimensione 382.26 kB
Formato Adobe PDF
382.26 kB Adobe PDF Visualizza/Apri
RECLAIM_Reverse_Engineering_Classification_Metrics.pdf

non disponibili

Tipologia: 2a Post-print versione editoriale / Version of Record
Licenza: Non Pubblico - Accesso privato/ristretto
Dimensione 775.43 kB
Formato Adobe PDF
775.43 kB Adobe PDF   Visualizza/Apri   Richiedi una copia
Pubblicazioni consigliate

Caricamento pubblicazioni consigliate

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11583/2971626