This paper takes the following view of #explainability: First train a model, and at test time, for a particular test instance, estimate the prediction Y, and then estimate the explanation E. Given this view, the paper uses the #PotentialOutcomes framework of #causality to understand the relationship between #prediction and explainability. Going from low to high performance of Y, the influence of Y on E is high, low, then high again.
#MachineLearning

https://arxiv.org/abs/2212.06925

#introduction

Hi,

I'm Joseph Bulbulia (Joe).

I teach #QuantitativePsychology at Victoria University of Wellington.

I serve on the leadership team of the New Zealand Attitudes and Values Study.

I supervise graduate students interested in #longitudinal methods for national-scale #paneldata.

My substantive research interests are in the psychology of religion, cultural evolution, moral psychology, well-being, and more recently #climatepsychology.

Big fan of the #potentialoutcomes framework for #causalinference.

Not a big fan of #prediction and #associations in #psychology.

I migrated to qoto.org because of its nice interface and science community.*

-- Joe

*Also, I think I can write things like

\[E(Y^1) - E(Y^0) \neq 0\]