Latent Dirichlet Allocation Nowadays I am trying to understand LDA. I have read a lot but something is missing. 
For instance, when I have a vector with three different variables (buy order, sell order, volume) respectively: [14,34,23]  And have over 12 Million records per each customer. 
I want to profile them by using LDA. How should I organize my data? What will be the analysing process? Furthermore, how will I interpret the results? I could not find any numerical (not topic modeling) examples. 
Thanks in advance..
Teitia 
 A: even if my answer is months after you posted your question, hope it can help others,
I have also recently started to learn about LDA, there is this video that can help Here, by the author of the paper.
and if you are using R, you can use the package LDA.
If I got you question right, you want to know, for each cluster what are the sell and buy orders that a repeated a lot? or you want to know what type of products he buys a lot?
we suppose you want to know the list of the products, and suppose he did 1 million purchase and for each purchase he has a list of elements.
You can first create vocabulary list, which is just a list that contains all the possible elements names (you can do it as the union of names of the elements in your 1 million purchase)
Then create a list of 1 million matrix, each list represent a purchase and in each list and the frequency of appearance, write the index of the element in the vocabulary (if using LDA package, it should be 0-based).
And they for example you say you want to get 5 categories of the elements that occur together.
Here is a sample code in R (I didn't test it, but it should work)
elem_vocab<-c("Elem1","Elem2","Elem3","Elem4","Elem5","Elem6","Elem7")
purchases<-list();
purchases[[1]]<-t(matrix(c( c(0,1,3,4), c(2,1,1,4) ),nrow=4,ncol=2))
purchases[[2]]<-t(matrix(c( c(0,2,3,5), c(1,6,1,5) ),nrow=4,ncol=2))
purchases[[3]]<-t(matrix(c( c(2,1,3,4), c(4,1,9,3) ),nrow=4,ncol=2))
purchases[[4]]<-t(matrix(c( c(6,1,5,4), c(3,10,3,4) ),nrow=4,ncol=2))

library(lda);
ldaRes<-lda.collapsed.gibbs.sampler(document=purchases,vocab=elem_vocab,
       K=2,num.iterations=200,alpha=0.1,eta=0.1,compute.log.likelihood=T,burnin=100)

topElem<-top.topic.words(ldaRes$topics,5,by.score=T)

Hope that helps
