explainableAI_tutorial
Preproccessing file: ESIP_preprocessing.ipynb
This file describes some of the basic steps in preproccessing data. The file called minimal_preprocessing.csv contains data that was already compiled by me from the CRMS stations. However, the only preprocessing done before what is shown in the ESIP_preprocessing.ipynb file is concatenation of dataframes and averaging using the .groupby method of pandas.
Machine Learning Tutorial: ESIP_tut_1_23_draft5.ipynb
Here is the whole tutorial on expainable AI. A summary is provided above for those interested in skipping to specific section of the tutorial. The structure of our tutorial follows below.
Motivation
Here we outline machine learning in the earth sciences and our specific use case. Advantages and disadvantages of machine learning are highlighted as well as the definition of explainable AI.
Machine Learning Tools
This is an expansive section in our tutorial. We begin by explaining linear regression, then explaining Bayesian Linear Regression (BLR) since this is a model we will implement in the tutorial. We then introduce the machine learning explainability python package called SHAP on the BLR.
Next, we introduce the Gaussian Process Regression Model. We implement this model using the scikit learn python library and conduct the same SHAP analysis on the GPR model to compare against the BLR SHAP analysis.
Data Preparation
Data Preparation is conducted primarily in the ESIP_preprocessing.ipynb file. Other preparations are done in the "Data Descriptions" portion of the tutorial and in scaling the the features.
Hyperparameter Tuning
The only example of hyperparameter tuning arises in the proof that the Regularized linear regression from both a Bayesian and Frequentist framework find the same solution. Here, we using hyperparamter tuing to find the optimal value for
Model training
We first train and implement our Bayesian Linear Regression model, analyze the results, then train the Gaussian Process Regression model. We do this using the scikit-learn python library.
Performance Evlaution
Performance Evaluation for both the Bayesian Linear Regression and the Gaussian Process Regression is done using repeated k-fold cross validation. For the Bayesian Linear Regression we use our own repeated k-fold cross validation and plotting function, while for the GPR we use an easy to implement repeated k-fold cross validation method from scikit learn.
Conclusion
We conduct a machine learning experiment in which we test if a nonlinear model, the Gaussian Process Regression, will work as a better approximator for accretion than a baseline linear model, the Bayesian Lienar Regression.