Your intuitive guide to interpret SHAP's beeswarm plot

The SHAP beeswarm plot is a powerful tool for interpreting machine learning models, but it can be a bit intimidating at first glance. The insights that you can drive from this plot is worth cracking its complexity, as it has become one of the standard steps after modeling.

So, why bother?

A SHAP beeswarm plot will help you:

1. Identify the most important features in a model: SHAP plot can show you which features have the biggest impact on the model's predictions. This can help you to understand how the model works and to identify any features that are redundant or unnecessary
2. Understand how individual features affect predictions: SHAP plot can show you how the predicted value of an observation changes as you vary the value of a single feature. This can help you to understand how the model is using each feature to make predictions.

So, dear reader, let's crack this plot's complexity together

Cracking the SHAP's beeswarm plot in 3 simple steps

To understand the SHAP's beeswarm plot, follow those simple 3 steps:

.

1. Review your model's target variable

End of the day, this plot represents the effect of input features on the final target variable prediction.

In our case, we trained a model on Stackoverflow's 2023 survey data to use the used languages and technologies used to predict if the role is a Data Scientist or a Data Engineer.

As the Data Scientist and Data Engineers are mutually exclusive in this set (i.e. you can be either one or another), I have decided to use only one target variable of both. And for some reason - AKA unconscious bias - I chose to model the target variable as the Data Scientist.

What does that mean:

• Data Scientist: The higher the predicted value is, the more likely for the prediction to be a Data Scientist
• Data Engineer: The lower the predicted value is, the more likely for the prediction to be a Data Engineer

.

.

2. Understand the SHAP's main axis

As with most plots of this world, the SHAP beeswarm plot also has an X and a Y axis.

I. X-axis: SHAP values:

Let's get started by how those values are calculated:

To intuitively explain those values, those values represent how did this input feature impact the final prediction. Those values have two components, the magnitude and the direction

• Magnitude: The higher the magnitude, the higher the impact of this feature on the prediction.
• Direction / sign: As some of the values are negative and some are positive, let's go back to our modeled target variable to understand this better. Remember that we have designed the model to predict a higher value for Data Science prediction and a lower value for a Data Engineer prediction? This is exactly what is meant by the direction / sign.

.

II. Y-axis: Input features

That's a simple one. Your input features are now listed on the y-axis and sorted according to their importance.

.

.

3. Crack a single feature's beeswarm sub-plot

For each of the input features, you will find a subplot that is called a beeswarm color-coded by the input feature value. To understand this better, let's split them into two parts:

• Beeswarm plot: a beeswarm is a plot that displays individual data points in a way that they don't overlap. That means that the width at a certain point represents the density.
• Color-coding: The colors represent the feature values. In our case, our features were binary, where 1 (red) represents that the technology is being used, and 0 (blue) represents that the technology is not being used. .

Let's put the pieces together to understand what's happening with our top feature: Scikit-learn:

You see that the red points (i.e. when Sklearn is used) have a positive SHAP values (i.e. prediction tends to be towards a Data Scientist)

.

Another example? Let's inspect Apache Kafka

You see that the red points (i.e. when Kafka is used) have a negative SHAP values (i.e. prediction tends to be towards a Data Engineer)

.

Yes, my dear reader, it is as simple as that :-)

.

Want to learn more?

Play around: A stand-alone notebook that I have developed for this illustration is now publicly available on Kaggle: https://www.kaggle.com/code/deenagergis/20231022-shap-beeswarm-illustration. Clone and play around!

Read more: For a more comprehensive explanation about the SHAP values, refer to Christoph Molnar's publicly available book https://christophm.github.io/interpretable-ml-book/shap.html