Note distinct y-axis scales for Terra and Aqua datasets. SHAP values showing the contribution of the MAIAC CWV to predictions of CWV measurement error shown across the time period of the study.The LOESS (locally estimated scatterplot smoothing) curve is overlaid in red. The color represents the MAIAC CWV for each observation (purple high, yellow low). SHAP values showing the contribution of the time trend to predictions.GridExtra::grid.arrange(g1, g2, ncol = 2) G2 <- (data_long = shap_long, x = 'dayint', y = 'Column_WV', color_feature = 'Column_WV') + ggtitle("(B) SHAP values of CWV vs. g1 <- (data_long = shap_long, x = 'dayint', y = 'dayint', color_feature = 'Column_WV') + ggtitle("(A) SHAP values of Time trend vs. Again, each dot is a station-day observation. It plots the SHAP values against the feature values for each variable. The SHAP scores (SHAP.Fever, SHAP.Cough) for model m1 and m2: In short, the order/structure of how the tree is built doesn’t matter for SHAP, but matters for Gain, and the mean absolute SHAP is the same (20 vs. 20). Xgb.importance(model = m1) # Feature Gain Cover Frequency Notice below the feature importance from xgb.importance were flipped. Use the dataset of Model A above as a simple example, which feature goes first into the dataset generates opposite feature importance by Gain: whichever goes later (lower in the tree) gets more credit. When we modify the model to make a feature more important, the feature importance should increase. Shap_dataĪnd why feature importance by Gain is inconsistentĬonsistency means it is legitimate to compare feature importance across different models. Pred_mod <- predict(mod, dataX, ntreelimit = 10) Shap_data <- copy(shap_values$shap_score) This is the case in this example, but not so if you are running e.g. 5-fold cross-validation. I.e., the explanation’s attribution values sum up to the model output (last column in the table below). As in the following table of SHAP values, rowSum equals the output predict(xgb_mod). The sum of each row’s SHAP values (plus the BIAS column, which is like an intercept) is the predicted model output. The SHAP values dataset ( shap_values$shap_score) has the same dimension (10148,9) as the dataset of the independent variables (10148,9) fit into the xgboost model.
SHAP values are calculated for each cell in the training dataset. Shap_values$mean_shap_score # dayint Column_WV AOT_Uncertainty dist_water_km aod Shap_values <- shap.values(xgb_model = mod, X_train = dataX)
# To return the SHAP values and ranked features by mean|SHAP| Verbose = FALSE, nthread = parallel::detectCores() - 2, Param_list <- list(objective = "reg:squarederror", # For regression Library("SHAPforxgboost") library("ggplot2") library("xgboost") Local explanation # run the model with built-in data
0 kommentar(er)
