如何移除shapviz生成的SHAP瀑布图Other Features并调整E[f(x)]位置

1. 移除“other features”并保留5个特征的瀑布图 #

默认sv_waterfall在设置max_display=5时会自动添加“other features”行汇总剩余特征。要移除它，需要过滤绘图数据并更新y轴刻度：

library(shapviz)
library(ggplot2)

# 生成原始瀑布图
g <- sv_waterfall(
  shap, row_id = 25, 
  max_display = 5, fill_colors = c("#e0115f", "#0096ff"), annotation_size = 5
)

# 过滤掉other features行
g$data <- g$data[!g$data$feature %in% "other features", ]

# 更新y轴的因子范围和刻度，确保只显示5个特征
y_scale <- g$scales$scales[[2]] # 定位y轴刻度对象
y_scale$breaks <- y_scale$breaks[!y_scale$breaks %in% "other features"]
y_scale$limits <- y_scale$limits[!y_scale$limits %in% "other features"]

# 输出定制后的图
print(g)

2. 调整max_display=10时E[f(x)]的位置 #

当显示10个特征时，E[f(x)]的注释位置可通过修改对应图层的x坐标调整：

# 生成10个特征的瀑布图
g_10 <- sv_waterfall(
  shap, row_id = 25, 
  max_display = 10, fill_colors = c("#e0115f", "#0096ff"), annotation_size = 5
)

# 定位E[f(x)]注释图层（通常为layers[[4]]，可通过str(g_10$layers)确认）
# 调整x坐标为基础值位置，或自定义数值（比如shap$baseline - 0.5）
g_10$layers[[4]]$data$x <- shap$baseline

# 输出调整后的图
print(g_10)

进阶：完全手动构造瀑布图（更灵活的控制） #

如果需要更精细的定制，可手动计算SHAP数据并使用ggplot直接绘制，彻底避免默认逻辑的限制：

# 提取目标行的SHAP值与特征名
shap_row <- shap$S[25, ]
feature_names <- colnames(shap$S)

# 按SHAP绝对值排序，取前5个特征
top5_idx <- order(-abs(shap_row))[1:5]
top5_shap <- shap_row[top5_idx]
top5_features <- feature_names[top5_idx]

# 构造瀑布图数据：基础值+特征SHAP+最终预测值
baseline <- shap$baseline
final_pred <- baseline + sum(top5_shap)

waterfall_data <- data.frame(
  feature = c("E[f(x)]", rev(top5_features), "f(x)"),
  value = c(baseline, rev(top5_shap), final_pred),
  cumulative = c(baseline, baseline + cumsum(rev(top5_shap)), final_pred),
  type = c("baseline", rep("feature", 5), "final")
)

# 绘制自定义瀑布图
ggplot(waterfall_data, aes(x = cumulative, y = reorder(feature, -row_number()), fill = type)) +
  geom_bar(stat = "identity") +
  geom_text(
    aes(label = round(value, 2)), 
    hjust = ifelse(waterfall_data$value >= 0, 1.1, -0.1), 
    size = 5
  ) +
  scale_fill_manual(
    values = c("baseline" = "gray", "feature" = c("#e0115f", "#0096ff"), "final" = "black")
  ) +
  labs(x = "SHAP Value", y = "Feature") +
  theme_minimal()

内容的提问来源于stack exchange，提问作者Ornella Scardua