Feature importance in machine learning models: A fuzzy information fusion approach

• DOI: 10.1016/j.neucom.2022.09.053
• Publication Date: 2022-09-13T06:35:32Z
• Authors: Divish Rengasamy; Jimiama M. Mase; Aayush Kumar; Benjamin Rothwell; Mercedes Torres T...; Morgan R. Alexander; David A. Winkler; Grazziela P. Figu...
• Keywords: Feature (linguistics); Representation; Feature Learning

With the widespread use of machine learning to support decision-making, it is increasingly important verify and understand reasons why a particular output produced. Although post-training feature importance approaches assist this interpretation, there an overall lack consensus regarding how should be quantified, making explanations model predictions unreliable. In addition, many these depend on specific approach employed subset data used when calculating importance. A possible solution improve reliability combine results from multiple quantifiers different coupled with re-sampling. Current state-of-the-art ensemble fusion uses crisp techniques fuse approaches. There is, however, significant loss information as are not context-aware reduce several single output. More importantly, their representation "importance" coefficients may difficult comprehend by end-users decision makers. Here we show fuzzy methods can overcome some limitations features easily understandable.