de Boer, Jens (2026) An empirical comparison of SHAP and LIME consistency for XGBoost-based extreme precipitation prediction in the Netherlands. Bachelor thesis, Data Science and Society (DSS).
|
PDF
BA4939255JLde-Boer.pdf Download (913kB) | Preview |
Abstract
Post-hoc explanation methods are commonly used to interpret machine learning models, but people often use them without checking if the explanations make sense or are reliable. This thesis looks at how SHAP and LIME fare when explaining an XGBoost classifier. The classifier is based on fifteen years of hourly ERA5 data and predicts extreme precipitation (above the 99th percentile) in De Bilt, the Netherlands. It focuses on whether the methods give consistent results, compared to each other. To test this, the study uses top-ten overlap and Spearman rank correlation to evaluate explanations for four different situations from the confusion matrix. SHAP and LIME end up agreeing a lot about the top features, mainly one-hour-lagged precipitation, though they differ once you get past the top five rankings.
| Item Type: | Thesis (Bachelor) |
|---|---|
| Name supervisor: | Schauble, J.K. |
| Date Deposited: | 10 Jun 2026 13:06 |
| Last Modified: | 10 Jun 2026 13:06 |
| URI: | https://campus-fryslan.studenttheses.ub.rug.nl/id/eprint/832 |
Actions (login required)
 |
View Item |