Project code for the Interpretable Machine Learning (IML) Seminar using the Ames Housing Prices dataset.
The dataset for this project can be found here in our repository, or be downloaded directly from Kaggle.
If you want to recreate the analysis set the working directory to the project folder. Then you can run the following files in this order:
| File | Content |
|---|---|
| cleaning_train.R | Data cleaning and pre-processing. Called by simple_models.R and xgboost_train.R to get the ready-to-use dataset (see section 1 in the report). |
| simple_models.R | Implementation and performance results of our 3 baseline (interpretable) models (see section 2 in the report). |
| xgboost_train.R | Implementation and performance results of our black box model - XGBoost. Saves the model for future use for our IML methods (see section 2 in the report). |
| outlier_detection.R | Analyze the dataset for presence of outliers (see section 3 in the report). |
| outlier_analysis.R | More in-depth look at the detected outliers (see section 3 in the report). |
| feature_importance.R | Implementation of feature importance algorithms for XGBoost and Linear Regression (see section 4.1 in the report). |
| specific_hypotheses.R | Analysis of hypotheses related to specific features (see section 4.2 in the report) |
| specific_points.R | Analysis of hypotheses related to specific points (see section 4.3 in the report) |
Files also contain extensive commentary for better code flow comprehension.
The following module versions (or above) are required to reproduce the code:
R = 4.0.3
dplyr = 1.0.2
plyr = 1.8.6
tidyverse = 1.3.0
readr = 1.4.0
ggplot2 = 3.3.2
corrplot = 0.84
caret = 6.0.86
Hmisc = 4.4.2
Metrics = 0.1.4
viridis = 0.5.1
partykit = 1.2.11
variables = 1.0.3
forcats = 0.5.0
data.table = 1.14.0
xgboost = 1.2.0.1
rpart = 4.1.15
rpart.plot = 3.0.9
glmnet = 4.0.2
cmaes = 1.0.11
mlr = 2.18.0
mlrCPO = 0.3.7
iml = 0.10.1
SHAPforxgboost = 0.1.0
featureImportance = 0.9
Require a manual installation from github! The github repository can be found here. The necessary packages are counterfactuals and the corresponding version of the iml package, which is different from the one listed above. We manually call the respective packages in our code (see specific_points.R).
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent