我有过去 3 年访问商店的数千名客户的数据。对于每个客户，我有：
Customer_Id | Year_*_Store 1 2010_A 1 2011_B 1 2012_C 2 2010_A 2 2011_B 2 2012_D
我想要的是以下数据结构，以便用河图（又名桑基图）可视化客户行为的演变
例如，2010 年第一次光顾 A 店的 2 位顾客，2011 年第一次光顾了 B 店：
SOURCE | TARGET | NB_CUSTOMERS
2010_A 2011_B 2
2011_B 2012_C 1
2011_B 2012_D 1
我不想要像 2010_A 和 2012_D 这样不连续的两年之间的链接
我会用 dplyr 来做这个（更快）
df<-read.table(header=T,text="Customer_Id Year_Store
1 2010_A
1 2011_B
1 2012_C
2 2010_A
2 2011_B
2 2012_D")
require(dplyr) # for aggregation
require(riverplot) # for Sankey
targets<-
group_by(df,Customer_Id) %.% # group by Customer
mutate(source=Year_Store,target=c(as.character(Year_Store)[-1],NA)) %.% # add a lag to show the shift
filter(!is.na(target)) %.% # filter out empty edges
regroup(list("source","target")) %.% # regroup by source & target
summarise(len=length(Customer_Id)) %.% # count customers for relationship
mutate(step=as.integer(substr(target,1,4))-as.integer(substr(source,1,4))) %.% # add a step to show how many years
filter(step==1) # filter out relationships for non consec years
topnodes <- c(as.character(unique(df$Year_Store))) # unique nodes
nodes <- data.frame( ID=topnodes, # IDs
x=as.numeric(substr(topnodes,1,4)), # x value for plot
col= rainbow(length(topnodes)), # color each different
labels= topnodes, # labels
stringsAsFactors= FALSE )
edges<- # create list of list
lapply(unique(targets$source),function(x){
l<-as.list(filter(targets,source==x)$len) # targets per source
names(l)<-filter(targets,source==x)$target # name of target
l
})
names(edges)<-unique(targets$source) # name top level nodes
r <- makeRiver( nodes, edges) # make the River
plot( r ) # plot it!
*请注意，列名中不能有 a （请参阅 参考资料?make.names）。这是一个基本的方法：
拆分Year_store为两个单独的列Year并Store在您的数据框中；目前它包含两种完全不同的数据，您实际上需要分别处理它们。
创建一个NextYear列，定义为Year + 1
做一NextStore列，指定店铺编码匹配Customer_Id且与Year本行相同NextYear，指定NA是否没有客户次年来店的记录，不符合要求则报错规格（对于第二年首先访问的商店模棱两可）。
NextStore去掉is中的任何行NA，并将NextYear和NextStore列组合成一NextYear_NextStore列。
Year_store通过和NextYear_NextStore列总结您的数据框，例如ddply在plyr包中使用。
一些样本数据：
# same example data as question
customer.df <- data.frame(Customer_Id = c(1, 1, 1, 2, 2, 2),
Year_Store = c("2010_A", "2011_B", "2012_C", "2010_A", "2011_B", "2012_D"),
stringsAsFactors = FALSE)
# alternative data should throw error, customer 2 is inconsistent in 2011
badCustomer.df <- data.frame(Customer_Id = c(1, 1, 1, 2, 2, 2),
Year_Store = c("2010_A", "2011_B", "2012_C", "2010_A", "2011_B", "2011_D"),
stringsAsFactors = FALSE)
和一个实现：
require(plyr)
splitYearStore <- function(df) {
df$Year <- as.numeric(substring(df$Year_Store, 1, 4))
df$Store <- as.character(substring(df$Year_Store, 6))
return(df)
}
findNextStore <- function(df, matchCust, matchYear) {
matchingStore <- with(df,
df[Customer_Id == matchCust & Year == matchYear, "Store"])
if (length(matchingStore) == 0) {
return(NA)
} else if (length(matchingStore) > 1) {
errorString <- paste("Inconsistent store results for customer",
matchCust, "in year", matchYear)
stop(errorString)
} else {
return(matchingStore)
}
}
tabulateTransitions <- function(df) {
df <- splitYearStore(df)
df$NextYear <- df$Year + 1
df$NextStore <- mapply(findNextStore, matchCust = df$Customer_Id,
matchYear = df$NextYear, MoreArgs = list(df = df))
df$NextYear_NextStore <- with(df, paste(NextYear, NextStore, sep = "_"))
df <- df[!is.na(df$NextStore),]
df <- ddply(df, .(Source = Year_Store, Target = NextYear_NextStore),
summarise, No_Customers = length(Customer_Id))
return(df)
}
结果：
> tabulateTransitions(customer.df)
Source Target No_Customers
1 2010_A 2011_B 2
2 2011_B 2012_C 1
3 2011_B 2012_D 1
> tabulateTransitions(badCustomer.df)
Error in function (df, matchCust, matchYear) :
Inconsistent store results for customer 2 in year 2011
没有尝试优化；如果您的数据集很大，那么也许您应该研究一个data.table解决方案。