Explainable AI with ICE ( Individual Conditional Expectation Plots )

In this article, we will go through what is ICE Plots and how we can use them to explain the Machine Learning Algorithm

Last blog I wrote about Explainable AI with PDP (Partial Dependence Plot), this next in a series that deals with another method for explaining ML algorithms.

If you hate theory and want to play with the code here is Google Colab for you.

For others, interested in how Individual Conditional Expectation Plots work read the entire story.

Introduction

Partial Dependence Plots have a serious problem of showing a line close to flat indicating no significant change in outcome value but in reality, the data can be of equal values with opposite signs. To address this issue, we have Individual Conditional Expectation (ICE) plots. They were introduced by researchers from The Wharton School in 2014 by building upon the work done by Friedman. They also intend to solve the problem by showing interactions between the variables which is not possible to a great extent in PDPs.

Definition

ICE plots are built on top of PDPs, they disaggregate the averaged data thus providing a chance to inspect the effect of the predictor variable at each value level while keeping the values of other predictor variables constant. A basic ICE plot shows how varying the feature value for an instance affects the predictive outcome by of course keeping other feature values constant. It can be cumbersome at times to analyze all the data points at once, but it also provides us a way to plot only one single point.

Let us again consider the same example of a class where the academic advisor still wants to analyze how students have done in the exam. But now he has a new problem, he wants to understand how each student has performed and does not want any class average. For the same subject ‘ExAI’, he can take a student’s scores from all subjects and keep replacing the student’s score with that of other students to understand how well he could have done marginally with the scores that other students might’ve gotten.

Consider N observations of {(XSi, XCi)} starting with i=1. Unlike PDPs where aggregated plots are produced by keeping the C set of features constant, here the plots are produced for every observed value in set S by keeping the same features constant. A curve is plotted for every fixed value of XC against observed values of XS. ICE plots solve the problem of providing insights about the model at a granular level, thus unearthing the average effect of PDP.

The volume of curves in a general ICE plot can be overwhelming and also intricate to understand at times.

The ICE plots have a variety of plots that can be plotted to make the analysis of a model more interesting and also takes it to a deeper level when needed:

Centered ICE plot

A general ICE plot may be visually challenging to understand at times and also to differentiate between 2 curves originating from different points. To solve this problem, the curves can be centered such that they originate from a single point. In doing so, the difference between the curves can be easily spotted. The plotted plot is called the “c-ICE” plot It is observed that choosing the Centralpoint as the least value of outcome variables gives the best results.

Derivative ICE plot

These plots are useful to investigate the presence of any interactions and the direction of the change in the predictive variable with respect to a feature by estimating the partial derivative of the curve. These are called the “d-ICE” plots. The derivative plots would display homogeneous curves showing only the difference in the level of prediction, heterogeneous curves would be present in case of any interactions of this feature with other features. The derivative plots can be given as:

.

Where g’(XS) is the derivative of the curve with respect to XS

Pros

  • They are intuitive and easy to implement
  • They can help in decoding the interactions between variables
  • They can help analyze the model at a granularity of each instance

Cons

  • It can get cumbersome to derive insights from ICE plots at times
  • They can only display the effects of only one feature at a time

Implementation

We use the Pycebox library and generate ICE plots

Visualizations

ICE plot

In the above plot, we see multiple lines plotted. Each line corresponds to a row in our data. We can see that for some individuals BMI does not affect the charges. But for a few of them, a high BMI seems to increase the charges. Such interpretation can be very useful in showing people the repercussions of having a high BMI.

ICE plot with PDP line

The above plot shows the ICE plot along with the aggregation line shown in black. The aggregation line is the same as the PDP line. The PDP line shows that the overall effect of the age is not much, though the charges increase a small bit after a certain age.

Centered ICE plot

The Centered ICE plot is centering the curves at a certain point in the feature and displays only the differences in prediction so that it is easy to interpret.

There are other visualizations that one can play and try to learn more from the notebook — Link here.

References

  1. Molnar, Christoph. “Interpretable machine learning. A Guide for Making Black Box Models Explainable”, 2019. https://christophm.github.io/interpretable-ml-book/. (the images were taken from here)
  2. The Elements of Statistical Learning: Trevor Hastie, Robert Tibshirani and Jerome Friedman

Special thanks to Kartik Kumar and Varun Raj for their contribution to the blog.

Stay tuned for the next blog on the ALE (Accumulated Local Effects Plot) which is built on the shortcomings of the PDP and ICE plots.

--

--

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store