Cluster observations based on multiple variables - r

Question

I am looking for an r function to create clusters in my dataset based on two variables (hope "cluster" is the right name for what I want to do) . Every two observations with the same value for variable_1 OR variable_2, should be in the same cluster. In the brief exemple that follows, I aggregating dataframe df based on variable_1 and variable_2.
df <- data.frame(variable_1=c("a","a","b","b","c","c","d","d","e","e"),variable_2=c("g1","g2","g1","g3","g2","g4","g4","g6","g7","g8"),value=rnorm(10))
df$clusters <- some_function_to_create_clusters(df[,c("variable_1","variable_2")])
The result should look like:
df$clusters <- c("clu1","clu1","clu1","clu1","clu1","clu1","clu1","clu1","clu2","clu2")
df
Notice that the first cluster contains everyone with variable_1 equals "a", "b", "c", or "d": "a" and "b" are merged together because they share "g1" (lines 1 and 3); "a" and "c" are merged because they share "g2" (lines 2 and 5); and "c" and "d" are merged because they share "g4" (lines 6 and 7).
Finally, in the last cluster there are only observations with variable_1=="e", because they don't share variable_2 with anyone.
Just to clarify what I intend to do, I will explain my problem set a little better. I’m pairing counties with nearby Touristic Attractions. Different counties are surround by different Touristic Attraction (TA), and there are many TA around the same county. But this "touristic-clusters" of counties and TAs are sparsely spread in the country. Notice that some far away counties might be inside the same cluster due to a “chain” effect of county-Touristic Attraction connections. So I want to find those “clusters” based on the id of the county and of the Touristic Attraction.
This seems simple, but I can't figure it out how to implement.
Thanks a lot

Answer 1

igraph solution
Disclaimer: I am completely new to igraph, so there's probably a better solution to this problem. However this seems to work.
With the igraph package we can cluster the data using the graph_from_data_frame() function, and then extract the clusters with components. You get the added advantage of being able to visualise the clusters.
library(igraph)
graph <- graph_from_data_frame(df[, 1:2], directed = FALSE)
cmp <- components(graph)$membership
df$cluster <- cmp[df$variable_1]
plot(graph)
Wrapping it up into a function
If you wanted to wrap it up as a function, something like this works:
find_clusters <- function(x, y) {
edges <- data.frame(from = x, to = y)
graph <- igraph::graph_from_data_frame(edges, directed = FALSE)
cmp <- igraph::components(graph)$membership
return(cmp[x])
}
Using the additional example you posted as a comment above, we thus have the following workflow:
library(dplyr)
df <- data.frame(
variable_1 = c("a", "a", "b", "b", "c", "c", "d", "d", "e", "e", "f", "f"),
variable_2 = c( "g1", "g2", "g1", "g3", "g2", "g4", "g4", "g6", "g7", "g8", "g9", "g12"),
value = rnorm(12)
)
df %>%
mutate(cluster = find_clusters(variable_1, variable_2))
# variable_1 variable_2 value cluster
# 1 a g1 -0.03410073 1
# 2 a g2 0.51261548 1
# 3 b g1 0.06470451 1
# 4 b g3 -1.97228101 1
# 5 c g2 -0.39751063 1
# 6 c g4 0.17761619 1
# 7 d g4 -0.13771207 1
# 8 d g6 -0.72183017 1
# 9 e g7 0.09012701 2
# 10 e g8 0.45763593 2
# 11 f g9 -0.83172613 3
# 12 f g12 2.83480352 3

Answer 2

So, I wrote a function to achieve what I need. It is ugly but it is working. If someone has a better/more efficient solution I would really appreciate it.
find_clusters <- function(original_df){
find_clus <- original_df
cluster_number <- 1
find_clus$cluster <- "cl"
i=1
for(i in 1:nrow(find_clus)){
if(nchar(find_clus$cluster[i])>2) next
aux <- lapply(original_df,function(x){ which(x==x[i])})%>% reshape2::melt()
idx <- aux$value %>%unique() %>%sort()
j = 1
while(j <= length(idx)){
aux <- lapply(original_df,function(x){ which(x==x[idx[j]])})%>% reshape2::melt()
idx <- c( idx, aux$value) %>%unique() %>% sort()
j <- j+1
}
find_clus$cluster[idx] <- paste0("cl",sprintf("%04d", cluster_number))
cluster_number<- cluster_number +1
}
return(find_clus$cluster)
}
So, to find the clusters one should write:
find_clusters(df[,c(1,2)])