2
$\begingroup$

I read a blog post about machine learning explainability - https://towardsdatascience.com/an-overview-of-model-explainability-in-modern-machine-learning-fc0f22c8c29a

It shows as an example this accumulated local effects plots (the y axis is bike rentals): accumulated local effects plot

In a previous example they showed the partial dependence plots on the same problem: partial dependence plot

According to the PDP plot until 25 degress as long as the temprature goes up the bike rentals goes up, so why are the values of the ALE there negative?

I thought that the ALE plot should be similar to derivative of the PDP (neglecting the fact that only close examples are taken), but it does not look that way.

Why does this plot look that way?

$\endgroup$

2 Answers 2

1
$\begingroup$

I found an answer here: https://christophm.github.io/interpretable-ml-book/ale.html

We accumulate the average effects across all intervals. The (uncentered) ALE of a feature value that lies, for example, in the third interval is the sum of the effects of the first, second and third intervals. The word Accumulated in ALE reflects this.

This effect is centered so that the mean effect is zero. The value of the ALE can be interpreted as the main effect of the feature at a certain value compared to the average prediction of the data. For example, an ALE estimate of -2 at $x_j=3$ means that when the j-th feature has value 3, then the prediction is lower by 2 compared to the average prediction.

$\endgroup$
0
$\begingroup$

As you answered yourself, the classic ALE centres data on zero, where zero means the mean value of the dataset. So, the PDP and ALE plots are quite similar once you shift the y-axis coordinates by approximately 4250 or so.

I find this not so intuitive, so in my new ale package in R, ALE values are centred on the median by default, which makes the plots more comparable to the PDP plots that you show above. However, it supports zero centring as well as an option, as well as mean centring.

$\endgroup$
1
  • $\begingroup$ I like the fact you actually implemented it better! $\endgroup$ Oct 22, 2023 at 18:05

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.