While working with some of the popular explainable AI libraries out there I kept running into the fact that it's quite a lot of work to set up explainers for your machine learning models. I figured: Why not ask the computer to come up with some useful explanations on its own?
I built a new tool called plsexplain that asks the computer: Please explain?!
You feed it a model and a dataset and it will show a beautiful interactive
dashboard that explains various aspects of your model.
Installing the dashboard
You can install the plsexplain tool by using the following command in your favorite terminal:
pip install plsexplain
Make sure you have scikit-learn and pandas installed as well. I haven't
included those in the package requirements just yet.
After installing the package, you can start creating your dashboard.
How the dashboard works
To create a explanation for your model, you can call the following command in your terminal:
plsexplain <model> <dataset>
When you press ENTER the application will start to build the dashboard for you. This usually takes up to 30 seconds depending on processor speed and the size of the dataset you provided.
While building the dashboard the program will show a nice progressbar.
Once the dashboard is ready, a new browser window is opened and you'll see a webpage similar to this:
You can immediately see the performance metrics and some information about your model. When you click on the "Model" page, you'll get a different view.
This page shows you the impact that each feature has on the outcome of your model. You can click on one of the features to zoom in and get more information about the impact of individual values of the feature on the outcome of the model.
Limitations and known issues
Right now, the dashboard functionality is limited to model-level explanations. I'm working on building prediction level explanations into the dashboard as well.
At the moment I'm recovering from an operation so it's slow going. However, I'll finish the work on the prediction-level explanations once I'm back on my feet.
The dashboard supports scikit-learn based models. And it requires a CSV file
with a header and comma as the separator for the fields.
Plans for the future
It's cool that you can have a basic model explanation with one command. To me it looks promising and I'm planning to add some more things to it:
- Support for tensorflow and pytorch: I'm planning on supporting these models so you can get insights into those types of models as well.
- Support for more dataset types: CSV files are pretty limited. Once I'm going to work on the deep-learning models, I'm also going to add support for images so that you can explain computer vision models as well as tabular models.
- Support for live monitoring: The offline use of this dashboard is nice, but I think it will be even more powerful when you add support for live monitoring. The idea is that you can send samples to the dashboard through an API and they will get explained as they come in.
The list is longer than I'm showing here. I have loads more ideas, but I'm also looking for feedback on this initial version of the dashboard.
Please let me know what you think!
Contributing
If you're interested in adding new features or fixing bugs, please head on over to Github: https://github.com/wmeints/plsexplain and submit your issues and pull requests there.