优化R代码执行速度：循环场景下data.table代码提速方案咨询

以下是几个能有效提升代码执行速度的优化方向，适配你在循环中动态调用列的场景：

1. 替换ifelse为逻辑索引直接求和 #

ifelse会生成额外的中间向量，改用逻辑向量直接索引目标列再求和，能减少内存开销和计算步骤：

ClusterName <- DT[, .(
  NumStart = sum(get(Num)[Start_End == Start_Contextual], na.rm = TRUE),
  NumEnd = sum(get(Num)[Start_End == End_Contextual], na.rm = TRUE),
  DenomStart = 1,
  DenomEnd = 1
), by = Cluster]

如果Start_End只有两类值，还可以用!复用逻辑条件，避免重复比较：

ClusterName <- DT[, {
  is_start = Start_End == Start_Contextual
  .(
    NumStart = sum(get(Num)[is_start], na.rm = TRUE),
    NumEnd = sum(get(Num)[!is_start], na.rm = TRUE),
    DenomStart = 1,
    DenomEnd = 1
  )
}, by = Cluster]

2. 移除eval(Cluster)，直接使用字符向量作为分组键 #

新版data.table支持直接将字符向量传入by参数，无需eval()解析，减少运行时的解析开销：

# 直接用by=Cluster代替by=eval(Cluster)
ClusterName <- DT[, {
  is_start = Start_End == Start_Contextual
  .(
    NumStart = sum(get(Num)[is_start], na.rm = TRUE),
    NumEnd = sum(get(Num)[!is_start], na.rm = TRUE),
    DenomStart = 1,
    DenomEnd = 1
  )
}, by = Cluster]

如果你的data.table版本较旧，可将字符转换为符号后传入：

Cluster_sym <- as.symbol(Cluster)
ClusterName <- DT[, {
  is_start = Start_End == Start_Contextual
  .(
    NumStart = sum(get(Num)[is_start], na.rm = TRUE),
    NumEnd = sum(get(Num)[!is_start], na.rm = TRUE),
    DenomStart = 1,
    DenomEnd = 1
  )
}, by = Cluster_sym]

3. 替换get(Num)为符号求值，避免运行时列查找 #

用rlang包的符号求值替代get()，能直接定位目标列，减少列名查找的开销（循环场景下收益更明显）：

library(rlang)
num_sym <- sym(Num)

ClusterName <- DT[, {
  is_start = Start_End == Start_Contextual
  .(
    NumStart = sum(!!num_sym[is_start], na.rm = TRUE),
    NumEnd = sum(!!num_sym[!is_start], na.rm = TRUE),
    DenomStart = 1,
    DenomEnd = 1
  )
}, by = Cluster]

4. 提前生成逻辑列，避免循环内重复条件判断 #

如果这段代码在循环中多次执行，可提前将Start_End转换为逻辑列，避免每次分组时重复比较字符串：

# 只在循环外执行一次
DT[, is_start := Start_End == Start_Contextual]

# 循环内的代码简化为：
ClusterName <- DT[, .(
  NumStart = sum(get(Num)[is_start], na.rm = TRUE),
  NumEnd = sum(get(Num)[!is_start], na.rm = TRUE),
  DenomStart = 1,
  DenomEnd = 1
), by = Cluster]

5. 预处理NA值，移除na.rm=TRUE的开销 #

如果目标列（Num对应的列）存在NA值，可提前将NA替换为0，这样求和时无需na.rm=TRUE，提升计算速度：

# 循环外预处理一次
DT[is.na(get(Num)), (Num) := 0]

# 循环内求和去掉na.rm参数
ClusterName <- DT[, .(
  NumStart = sum(get(Num)[is_start]),
  NumEnd = sum(get(Num)[!is_start]),
  DenomStart = 1,
  DenomEnd = 1
), by = Cluster]

内容的提问来源于stack exchange，提问作者paulbouu