LIME-research-journals

LIME Research

Repository for research journals for my research projects related to LIME. See below for descriptions of the contents of this repository.

Descriptions of the work done can be found in the journals. Beware of typos in the journals. Not all of them have been proof read. For easy viewing of the journals, use these links:

1. Objectives and Ideas
2. Information on Hamby Data and Models
3. Understanding LIME
4. Applying LIME to Hamby Data
5. Assessing LIME on Hamby Data
6. Applying and Assessing LIME on Iris Data
7. Notes on Readings
8. Computational Issues
9. LIME Applied to Logistic Regressions on Sine Data
10. Identifying rtrees Training Data
11. LIME Variability Across Hamby Data and Models
12. Gower Exponent Examples

Repository Contents

Descriptions of the materials in this repository are listed below.

code

Place for storing R code used in the research journals:

• app.R: code for the Shiny app for visualizing the LIME explanations associated with the bullet land comparisons
• helper_functions.R: R script with functions that I use in the journals

figures

Contains static images:

• laugel-lime-concept.png and laugel-locality-metric: figures from the Laugel et al. (2018) paper
• python-ridge1, python-ridge2, and python-ridge3: screen shots of the python LIME package showing the use of a ridge regression as the explainer

journals

Contains folders with research journals:

• 00-objectives_and_ideas: background of the research project, goals of the project, concerns with LIME, and ideas to try
• 01-hamby_data_and_models***: information on the Hamby data, cleaning of the training and testing data, visualizations of the data, and models fit to the data (rtrees and logistic regressions)
• 02-understanding_lime***: work done to understand the LIME algorithm
• 03-applying_lime***: applications of LIME to the models fit to the Hamby data
• 04-assessing_lime***: visual diagnostics for assessing the lime explanations
• 05-iris_comparison: applying LIME to a random forest fit to the iris data and assessing the explanations
• 06-literature_review: notes on papers relating to LIME
• 07-computation_issues***: documentation of some of the computation issues I ran into during the project
• 08-logistic_regression: examples of applying and diagnosing LIME with logistic regression models fit to the sine data
• 09-identifying_rtrees_training_data: examples where I retrain the rtrees model with the “current” version of the training data and compare the results
• 10-lime_hamby_variability: investigating variability between LIME explanations using diagnostic figures for random forest models trained on the Hamby bullet data

*** Note: We have realized that the training dataset for the rtrees model is not the one used in these files. See 09-identifying_rtrees_training_data for more details on the correct dataset. The correct data is used in the LIME diagnostics paper. For time reasons, the results in these journals have not been updated.

old-journals

Contains materials from my earliest research journals relating to the LIME project. It is likely that the code in these files will not run. However, the knit versions of the R markdown documents may still be useful, so I included them in the repository.

main

Files contained in the main folder of the repository:

• .gitignore: files to ignore
• notes from Heike.pdf: scans of notes written by Heike for this project
• README.md and README.Rmd: this README and the R markdown document used to generate this README