Neural Additive Models (NAMs) Agarwal et al. 2020 combine some of the expressivity of DNNs with the inherent intelligibility of generalized additive models. NAMs learn a linear combination of neural networks that each attend to a single input feature. These networks are trained jointly and can learn arbitrarily complex relationships between their input feature and the output.
In this visualization approach, feature pair heatmaps (2D-heatmaps) and their corresponding feature maps (shape functions) are visualized in a Dash app. The user can select between automatic filtering of the most useful heatmaps via variance of the heatmaps and manual selection of the heatmaps. The manual mode is also supported by permutation feature importance on the full NAM model.
As stated in the NAM paper, the shape function and the normalized data density are plotted on the same graph. The normalized data density is visualized in the form of red bars. The darker the shade of red, the more data there is in that region. This allows us to know when the model had adequate training data to learn appropriate shape functions.
This code was tested with Python 3.8
Create a virtual environment using this guideline: Installing packages using pip and virtual environments
Create a virtual environment using this guideline: Creating an environment with commands (conda)
Check your system, if it's CUDA capable or not, and use the corresponding version of PyTorch.
PyTorch versions with or without CUDA support.
If necessary, change the pip statement below with respect to your PyTorch version (torch).
After that, install the needed packages via this pip statement:
pip install numpy pandas torch torchmetrics scikit-learn plotly dash dash_daq dash-extensions pyautoguiOr download the requirements.txt from this repository and use this statement:
pip install -r requirements.txt- Download
app.pyfrom this repository. - Activate your virtual environment.
- Go to the directory where
app.pyis located. - Run the following command in the command line:
python app.pyAfter this step, a web browser window should be opened with the dash app.
As an input, datasets in the .csv format are allowed.
Drag & Drop the file into the marked area, or simply click on Select File.
Now you should see the head of the .csv file and the name of the file.
Important Notice!!!
The dataset has to be free of NaN values or other kinds of missing values --> it should be properly preprocessed!
If you can't preprocess the data properly, you can also just select the feature columns which are working --> which have no NaN values.
2023-02-17_16-32-23.mp4
A simple implementation of the Neural Additive Model by Agarwal et al. in PyTorch.
Neural Additive Models (Google Research)
Neural Additive Models: Interpretable Machine Learning with Neural Nets