This function is passed to the matrix, where each column is an element, and each row is a membership vector corresponding to the partition of elements in accordance with the clustering method. The elements (numbers) that make up each line have no meaning other than indicating ownership and are processed from line to line. The function returns the majority vote section. When consensus does not exist for an element, the section specified by the first line wins. This allows you to organize sections, for example, by decreasing modularity values.
consensus.final <- function(data){ output=list() for (i in 1:nrow(data)){ row=as.numeric(data[i,]) output.inner=list() for (j in 1:length(row)){ group=character() group=c(group,colnames(data)[which(row==row[j])]) output.inner[[j]]=group } output.inner=unique(output.inner) output[[i]]=output.inner } # gives the mode of the vector representing the number of groups found by each method consensus.n.comm=as.numeric(names(sort(table(unlist(lapply(output,length))),decreasing=TRUE))[1]) # removes the elements of the list that do not correspond to this consensus solution output=output[lapply(output,length)==consensus.n.comm] # 1) find intersection # 2) use majority vote for elements of each vector that are not part of the intersection group=list() for (i in 1:consensus.n.comm){ list.intersection=list() for (p in 1:length(output)){ list.intersection[[p]]=unlist(output[[p]][i]) } # candidate group i intersection=Reduce(intersect,list.intersection) group[[i]]=intersection # we need to reinforce that group for (p in 1:length(list.intersection)){ vector=setdiff(list.intersection[[p]],intersection) if (length(vector)>0){ for (j in 1:length(vector)){ counter=vector(length=length(list.intersection)) for (k in 1:length(list.intersection)){ counter[k]=vector[j]%in%list.intersection[[k]] } if(length(which(counter==TRUE))>=ceiling((length(counter)/2)+0.001)){ group[[i]]=c(group[[i]],vector[j]) } } } } } group=lapply(group,unique) # variables for which consensus has not been reached unclassified=setdiff(colnames(data),unlist(group)) if (length(unclassified)>0){ for (pp in 1:length(unclassified)){ temp=matrix(nrow=length(output),ncol=consensus.n.comm) for (i in 1:nrow(temp)){ for (j in 1:ncol(temp)){ temp[i,j]=unclassified[pp]%in%unlist(output[[i]][j]) } } # use the partition of the first method when no majority exists (this allows ordering of partitions by decreasing modularity values for instance) index.best=which(temp[1,]==TRUE) group[[index.best]]=c(group[[index.best]],unclassified[pp]) } } output=list(group=group,unclassified=unclassified) }
Some examples:
data=cbind(c(1,1,2,1,1,3),c(1,1,2,1,1,1),c(2,2,1,2,1,2)) colnames(data)=paste("item",1:3) rownames(data)=paste("method",1:6) data consensus.final(data)$group [[1]] [1] "item 1" "item 2" [[2]] [1] "item 3" data=cbind(c(1,1,1,1,1,3),c(1,1,1,1,1,1),c(1,1,1,2,1,2)) colnames(data)=paste("item",1:3) rownames(data)=paste("method",1:6) data consensus.final(data)$group [[1]] [1] "item 1" "item 2" "item 3" data=cbind(c(1,3,2,1),c(2,2,3,3),c(3,1,1,2)) colnames(data)=paste("item",1:3) rownames(data)=paste("method",1:4) data consensus.final(data)$group [[1]] [1] "item 1" [[2]] [1] "item 2" [[3]] [1] "item 3"