模型对比
R 中的树模型机器学习
Sandro Raabe
Data Scientist
动机
比较 AUC
# A tibble: 4 x 3
model .metric .estimate
1 decision_tree roc_auc <?>
2 bagged_trees roc_auc <?>
3 random_forest roc_auc <?>
4 boosted_trees roc_auc <?>
比较 ROC 曲线
合并预测值
bind_cols(decision_tree
)
# A tibble: 1,011 x 1
preds_tree
<dbl>
1 0.144
2 0.441
3 0.144
4 0.776
5 0.441
6 0.144
7 0.144
8 0.441
# ... 另有 1,003 行
合并预测值
bind_cols(decision_tree, bagged_trees
)
# A tibble: 1,011 x 2
preds_tree preds_bagging
<dbl> <dbl>
1 0.144 0.115
2 0.441 0.326
3 0.144 0.115
4 0.776 0.773
5 0.441 0.326
6 0.144 0.115
7 0.144 0.115
8 0.441 0.877
# ... 另有 1,003 行
合并预测值
bind_cols(decision_tree, bagged_trees, random_forest
)
# A tibble: 1,011 x 3
preds_tree preds_bagging preds_forest
<dbl> <dbl> <dbl>
1 0.144 0.115 0
2 0.441 0.326 0
3 0.144 0.115 0
4 0.776 0.773 0.286
5 0.441 0.326 0.15
6 0.144 0.115 0
7 0.144 0.115 0
8 0.441 0.877 0.7
# ... 另有 1,003 行
合并预测值
bind_cols(decision_tree, bagged_trees, random_forest, boosted_trees
)
# A tibble: 1,011 x 4
preds_tree preds_bagging preds_forest preds_boosting
<dbl> <dbl> <dbl> <dbl>
1 0.144 0.115 0 0.136
2 0.441 0.326 0 0.149
3 0.144 0.115 0 0.116
4 0.776 0.773 0.286 0.319
5 0.441 0.326 0.15 0.199
6 0.144 0.115 0 0.116
7 0.144 0.115 0 0.116
8 0.441 0.877 0.7 0.823
# ... 另有 1,003 行
合并预测值
bind_cols(decision_tree, bagged_trees, random_forest, boosted_trees,
customers_test %>% select(still_customer))
# A tibble: 1,011 x 5
preds_tree preds_bagging preds_forest preds_boosting still_customer
<dbl> <dbl> <dbl> <dbl> <fct>
1 0.144 0.115 0 0.136 no
2 0.441 0.326 0 0.149 no
3 0.144 0.115 0 0.116 no
4 0.776 0.773 0.286 0.319 yes
5 0.441 0.326 0.15 0.199 no
6 0.144 0.115 0 0.116 no
7 0.144 0.115 0 0.116 no
8 0.441 0.877 0.7 0.823 yes
# ... 另有 1,003 行
计算决策树 AUC
# 计算 AUC 指标
roc_auc(preds_combined, truth = still_customer, estimate = preds_tree)
# A tibble: 1 x 2
.metric .estimate
<chr> <dbl>
1 roc_auc 0.911
计算装袋树 AUC
# 计算 AUC 指标
roc_auc(preds_combined, truth = still_customer, estimate = preds_bagging)
# A tibble: 1 x 2
.metric .estimate
<chr> <dbl>
1 roc_auc 0.936
计算随机森林 AUC
# 计算 AUC 指标
roc_auc(preds_combined, truth = still_customer, estimate = preds_forest)
# A tibble: 1 x 2
.metric .estimate
<chr> <dbl>
1 roc_auc 0.974
计算提升法 AUC
# 计算 AUC 指标
roc_auc(preds_combined, truth = still_customer, estimate = preds_boosting)
# A tibble: 1 x 2
.metric .estimate
<chr> <dbl>
1 roc_auc 0.984
合并所有 AUC
# 合并各模型的 AUC
bind_rows(roc_auc(preds_combined, truth = still_customer, estimate = preds_tree),
roc_auc(preds_combined, truth = still_customer, estimate = preds_bagging),
roc_auc(preds_combined, truth = still_customer, estimate = preds_forest),
roc_auc(preds_combined, truth = still_customer, estimate = preds_boosting))
# A tibble: 4 x 2
.metric .estimate
<chr> <dbl>
1 roc_auc 0.911
2 roc_auc 0.936
3 roc_auc 0.974
4 roc_auc 0.984
合并所有 AUC
# 合并各模型的 AUC
bind_rows(decision_tree = roc_auc(preds_combined, truth = still_customer, preds_tree),
bagged_trees = roc_auc(preds_combined, truth = still_customer, preds_bagging),
random_forest = roc_auc(preds_combined, truth = still_customer, preds_forest),
boosted_trees = roc_auc(preds_combined, truth = still_customer, preds_boosting),
.id = "model")
# A tibble: 4 x 3
model .metric .estimate
<chr> <chr> <dbl>
1 decision_tree roc_auc 0.911
2 bagged_trees roc_auc 0.936
3 random_forest roc_auc 0.974
4 boosted_trees roc_auc 0.984
重整结果格式
# 将预测转为长格式 predictions_long <- tidyr::pivot_longer(preds_combined,cols = starts_with("preds_"),names_to = "model",values_to = "predictions")
# A tibble: 4,044 x 3
still_customer model predictions
<fct> <chr> <dbl>
1 no preds_tree 0.144
2 no preds_bagging 0.102
3 no preds_forest 0.0333
4 no preds_boosting 0.169
5 yes preds_tree 0.441
6 no preds_bagging 0.285
7 no preds_forest 0.36
8 no preds_boosting 0.184
# ... 另有 4,036 行
计算阈值指标
# 按模型分组 cutoffs <- predictions_long %>% group_by(model) %>%# 计算每个阈值的指标 roc_curve(truth = still_customer, estimate = predictions)
# A tibble: 668 x 4
# Groups: model [4]
model .threshold specificity sensitivity
<chr> <dbl> <dbl> <dbl>
1 preds_bagging -Inf 0 1
2 preds_bagging 0.0157 0 1
3 preds_bagging 0.0202 0.536 0.975
4 preds_bagging 0.0254 0.537 0.975
5 preds_bagging 0.0271 0.665 0.938
6 preds_bagging 0.0315 0.681 0.938
# ... 另有 662 行
绘制 ROC 曲线
# 转为可视化
autoplot(cutoffs)
来比较一下！
R 中的树模型机器学习
