求助：geom_smooth未呈现预期水平曲线（不愿调整span）

问题原因 #

你遇到的是loess局部加权回归的边缘效应：当第一个数据点与后续点间隔较大时，局部回归在边缘区域的权重分配会导致拟合曲线偏离真实值，出现异常峰值。由于你不想调整span参数（会影响其他数据的拟合效果），可以通过以下两种针对性方案解决：

方案一：自定义平滑逻辑，动态切换拟合方法 #

通过自定义函数判断每组数据的y值是否为常数，常数数据用线性回归拟合水平线，非常数数据保留默认loess方法：

# 自定义平滑方法
custom_smooth <- function(formula, data, ...) {
  if (length(unique(data$y)) == 1) {
    # 常数数据拟合水平线
    lm(formula, data = data, ...)
  } else {
    # 非常数数据使用默认loess
    loess(formula, data = data, ...)
  }
}

# 修改后的绘图代码
ggplot(test, aes(month_day, Sensibilite, label = Nom, color = Nom)) +
  geom_smooth(method = custom_smooth, se = FALSE, size=0.5) +
  geom_point(size =0.5)+ 
  geom_label_repel(aes(label = after_stat(helper_label(x, color, PANEL, "max"))),
                   stat = "smooth",
                   method = custom_smooth, # 同步指定自定义方法
                   segment.size = 0.1,
                   max.overlaps = Inf, size = 2, na.rm = TRUE, direction = "y", nudge_x = 30
  ) +
  theme_light()+
  theme(legend.position = "none")+ 
  scale_y_continuous(limits = c(0,6))+
  scale_x_date(expand = c(0.15, 0),name = "Mois",date_breaks = "1 month",date_labels = "%b")

方案二：调整loess的边界控制参数 #

无需切换方法，仅修改loess的surface参数为"direct"，限制拟合范围在数据点区间内，避免边缘外推：

ggplot(test, aes(month_day, Sensibilite, label = Nom, color = Nom)) +
  geom_smooth(method = "loess", 
              method.args = list(control = loess.control(surface = "direct")),
              se = FALSE, size=0.5) +
  geom_point(size =0.5)+ 
  geom_label_repel(aes(label = after_stat(helper_label(x, color, PANEL, "max"))),
                   stat = "smooth",
                   method = "loess",
                   method.args = list(control = loess.control(surface = "direct")),
                   segment.size = 0.1,
                   max.overlaps = Inf, size = 2, na.rm = TRUE, direction = "y", nudge_x = 30
  ) +
  theme_light()+
  theme(legend.position = "none")+ 
  scale_y_continuous(limits = c(0,6))+
  scale_x_date(expand = c(0.15, 0),name = "Mois",date_breaks = "1 month",date_labels = "%b")

效果说明 #

两种方案都能让常数数据的拟合曲线保持水平线，同时完全保留原span参数对其他非常数数据的拟合效果。

内容的提问来源于stack exchange，提问作者Mmerose