R语言SDC包ProtectTable()不收敛问题求解

我有一个包含4个多标签分类变量的数据集：两个变量的基数为4，一个为2，一个为3。为了支持单个变量内多个值为真的场景，我将这些变量独热编码为二进制矩阵，最终得到13个基数为2的变量。但此时使用sdc包的ProtectTable()函数时无法收敛。我曾尝试创建变量所有可能组合的联合变量，但披露控制未成功；未编码原始变量时函数运行正常。

onehotencoded <- df_ref %>%
  mutate(
    FQuarter_1 = if_else(FQuarter == "Q1", 1, 0),
    FQuarter_2 = if_else(FQuarter == "Q2", 1, 0),
    FQuarter_3 = if_else(FQuarter == "Q3", 1, 0),
    FQuarter_4 = if_else(FQuarter == "Q4", 1, 0),
    FYear_24 = if_else(FYear == "2024", 1, 0),
    FYear_23 = if_else(FYear == "2023", 1, 0),
    FYear_22 = if_else(FYear == "2022", 1, 0),
    FYear_21 = if_else(FYear == "2021", 1, 0),
    Working_model_FT = if_else(Working_model == "Full-time", 1, 0),
    Working_model_PT = if_else(Working_model == "Part-time", 1, 0),
    Gender_Male = if_else(Gender == "men", 1, 0),
    Gender_Female = if_else(Gender == "women", 1, 0),
    Gender_Diverse = if_else(Gender == "else", 1, 0)
  ) %>%
  select(-FQuarter, -FYear, -Working_model, -Gender)


df_ref <- onehotencoded


# Negative Threshold ------------------------------------------------------

negative_threshold_function <- function(df_ref, threshold) {
  # convert inputs to character type
  df_ref <- df_ref %>% mutate_all(as.character)
  
  # Create dimList object dynamically for all dimensions in df_ref (except the first column org_level_hierachy)
  dimList <- list()
  
  for (col_name in colnames(df_ref)) {
    dim_hierarchy <- sdcHierarchies::hier_create(root = "Total", nodes = unique(df_ref[[col_name]]))
    dim_data <- matrix(nrow = nrow(dim_hierarchy), ncol = 2)
    dim_data[,1] <- sapply(dim_hierarchy$level, function(x) paste(replicate(x, "@"), collapse = ""))
    dim_data[,2] <- dim_hierarchy$leaf
    colnames(dim_data) <- c("levels", "codes")
    dim_data <- as.data.frame(dim_data)
    dimList[[col_name]] <- dim_data
  }
  
  # Protect table, with secondary suppression
  df_aggregated_protected <- ProtectTable(df_ref,
                                          dimList = dimList,
                                          infoAsFrame = T,
                                          #method = "HITAS",
                                          method = "SIMPLEHEURISTIC_OLD",
                                          maxN = threshold,
                                          maxVars = 2,
                                          #maxVars = ncol(df_ref) - 1,
                                          fastSolution = TRUE,
                                          save = FALSE, # save tables in between
                                          timeLimit = 0.1, #so to say a timeout, set very low, trying to fasten computation
                                          verbose = TRUE) #"Total" means filter is not selected.
  
  #small adjustments: sdcStatusLegend, Suppression column
  df_aggregated_protectedFinal <- df_aggregated_protected[["data"]] %>% 
    mutate(sdcStatusLegend = case_when(
      sdcStatus == "u" ~ "cell is primary suppressed and needs to be protected",
      sdcStatus == "x" ~ "cell has been secondary suppressed",
      sdcStatus == "s" ~ "cell can be published",
      sdcStatus == "z" ~ "cell must not be suppressed"
    )) %>% 
    mutate(suppressed = case_when(
      is.na(suppressed) ~ "yes",
      !is.na(suppressed) ~ NA_character_
    )) %>% 
    rename(group_size = freq, suppress = suppressed) %>% 
    relocate(sdcStatus, .after = suppress)
  
  return(df_aggregated_protectedFinal)
}

runtime <- system.time({
  final_table <- negative_threshold_function(df_ref, threshold = 5)
})

1. 调整ProtectTable()参数解决收敛问题 #

独热编码后维度数量剧增（13个）导致计算量爆炸，当前参数设置过于严苛是无法收敛的核心原因：

• 增大timeLimit：将timeLimit从0.1调整为60甚至300，给算法足够时间完成计算。
• 关闭fastSolution：设置fastSolution = FALSE，牺牲部分速度换取更高的收敛概率。
• 切换抑制算法：尝试method = "HITAS"，该算法在高维度场景下比SIMPLEHEURISTIC_OLD更稳定。
• 调整maxVars：适当增大maxVars（比如设为3或4），允许算法考虑更多变量组合的抑制策略，避免陷入局部最优。

调整后的核心代码片段：

df_aggregated_protected <- ProtectTable(df_ref,
                                        dimList = dimList,
                                        infoAsFrame = T,
                                        method = "HITAS",
                                        maxN = threshold,
                                        maxVars = 3,
                                        fastSolution = FALSE,
                                        save = FALSE,
                                        timeLimit = 60,
                                        verbose = TRUE)

2. 避免独热编码，直接处理多标签变量 #

独热编码会大幅增加维度数，转而直接处理多标签变量的组合更高效：

• 生成标签组合变量：将每个原始多标签变量的取值转换为标签组合字符串（比如某观测同时属于Q1和Q2，FQuarter取值为"Q1+Q2"），保留原始的4个变量维度，而非拆分为13个二进制变量。
• 构建合理层级：使用sdcHierarchies::hier_create()为每个组合变量创建层级结构（比如将单个标签作为底层，组合标签作为上层），帮助ProtectTable()利用层级减少计算复杂度。

示例代码（以FQuarter为例生成组合）：

library(tidyr)
library(dplyr)

# 假设原始df_ref中FQuarter是多标签（用逗号分隔）
df_ref <- df_ref %>%
  mutate(FQuarter_comb = case_when(
    grepl("Q1", FQuarter) & grepl("Q2", FQuarter) ~ "Q1+Q2",
    grepl("Q1", FQuarter) & grepl("Q3", FQuarter) ~ "Q1+Q3",
    # 补充其他组合情况
    TRUE ~ FQuarter
  )) %>%
  select(-FQuarter)

3. 手动构建sdcProblem对象 #

如果上述方法仍不生效，可以手动构建问题对象，精准控制计算范围：

• 使用sdcTable::makeProblem()函数，先定义所有需要聚合的维度和频数阈值。
• 手动标记需要抑制的单元格，再调用sdcTable::protectTable()处理，减少不必要的计算开销。

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