I have a lottery style dataset we produce internally (example below). I am trying to figure out which numbers appear most frequently together. Example questions: What are the top 10 pair of numbers that appear most frequently together? What are the top 10 three numbers that appear most frequently together?

What methods/techniques would I need to use to answer these questions?

I was thinking clustering at first but a test run (ward) I performed, didn't return the results I was expecting.

enter image description here

  • 1
    $\begingroup$ You might be able to mine a closely related thread for useful ideas. Provided you have few columns (only five are shown in the example), it would take little to adapt the solutions. $\endgroup$
    – whuber
    Nov 12 '13 at 20:17
  • 2
    $\begingroup$ On the question* "which pairs of numbers appear most often together", I'd start with cross-tabulations. $\quad\,\,\,$ $\, $* the 'question' is actually a host of somewhat different questions, depending on things like whether the pair (Num1=3,Num2=5) is considered the same as (Num1=5,Num2=3) (does order matter?), or whether it's considered the same as (Num1=3,Num5=5) (do the column labels matter?) $\endgroup$
    – Glen_b
    Nov 12 '13 at 21:42
  • $\begingroup$ @whuber...thanks for the link. That sounds like something that could be modified to answer my question. And I am using r. But not sure I have the necessary skill level yet to adapt that r code. I'll give it shot though. $\endgroup$ Nov 12 '13 at 22:25
  • $\begingroup$ @Glen_b Great clarifying question. Order doesn't matter. The three examples you gave are all equivalent for the question I am asking. Also, the numbers are non-repeating for each row. $\endgroup$ Nov 12 '13 at 22:33
  • 1
    $\begingroup$ Then it sounds like you could collapse all the pairwise cross-tabulations into a single table (add the counts from all pairwise tables), then add the counts from that single table to its transpose (since order doesn't matter), and take the lower triangular portion of that combined table (since its symmetric). Then find the biggest counts in that table ... and that gives the pairs that occur together the most. (If that's useful to you let me know and I'll write it up as an answer) $\endgroup$
    – Glen_b
    Nov 12 '13 at 22:58

This question calls for a modification of the solution to a sequence counting problem: as noted in comments, it requests a cross-tabulation of co-occurrences of values.

I will illustrate a naive but effective modification with R code. First, let's introduce a small sample dataset to work with. It's in the usual matrix format, one case per row.

x <- matrix(c(3,5,7,10,13,
            ncol=5, byrow=TRUE)

This solution generates all possible combinations of $m$ items (per row) at a time and tabulates them:

m <- 3
x <- t(apply(x, 1, sort))
x0 <- apply(x, 1, combn, m=m)
y <- array(x0, c(m, length(x0)/(m*dim(x)[1]), dim(x)[1]))
ngrams <- apply(y, c(2,3), function(s) paste("(", paste(s, collapse=","), ")", sep=""))
z <- sort(table(as.factor(ngrams)), decreasing=TRUE)

The tabulation is in z, sorted by descending frequency. It is useful by itself or easily post-processed. Here are the first few entries of the example:

> head(z, 10)
 (3,5,10) (3,10,15)  (3,5,15)   (3,5,8) ... (8,10,15) 
        3         2         2         2 ...         2

How efficient is this? For $p$ columns there are $\binom{p}{m}$ combinations to work out, which grows as $O(p^m)$ for fixed $m$: that's pretty bad, so we are limited to relatively small numbers of columns. To get a sense of the timing, repeat the preceding with a small random matrix and time it. Let's stick with values between $1$ and $20,$ say:

n.col <- 8       # Number of columns
n.cases <- 10^3  # Number of rows
x <- matrix(sample.int(20, size=n.col*n.cases, replace=TRUE), ncol=n.col)

The operation took two seconds to tabulate all $m=3$-combinations for $1000$ rows and $8$ columns. (It can go an order of magnitude faster by encoding the combinations numerically rather than as strings; this is limited to cases where $\binom{p}{m}$ is small enough to be represented exactly as an integer or float, limiting it to approximately $10^{16}$.) It scales linearly with the number of rows. (Increasing the number of possible values from $20$ to $20,000$ only slightly lengthened the time.) If that suggests overly long times to process a particular dataset, then a more sophisticated approach will be needed, perhaps utilizing results for very small $m$ to limit the higher-order combinations that are computed and counted.

  • $\begingroup$ @ whuber. Thanks for your posted Answer. Why is the output of z showing a count of 5 for (3,5,10)? Based on the original matrix I was expecting to see a count of 3. This would come from rows x[c(1:2,6),]. If I enter m<-5 I would expect to see in the output of z (3,5,8,10,15) with a count of 2 and and 5 additional groups each with a count of 1. $\endgroup$ Nov 14 '13 at 18:31
  • $\begingroup$ Sorry: although I checked that, I misread the output and did not catch the discrepancy. There was a typo in the code where "2" appeared but "m" should have (and R issued no warnings about a mismatch between the sizes of arrays x0 and y; it merely "recycled" the elements). I fixed that typo and pasted the new output. You might find it informative (and reassuring) to run the code and inspect the array y, which contains all the $m$-tuples for each row in x. As another check, sum(z) should equal $\binom{p}{m}$ times the number of rows in $x$: $\binom{5}{3}\times 6=60$ in the example. $\endgroup$
    – whuber
    Nov 14 '13 at 18:41
  • 1
    $\begingroup$ np. I was actually just inspecting y as you were writing your answer and thinking "I was not expecting to see 15 columns for m<-3..Should be choose(5,3) =10". But your edit fixed that. This is way cool! Thank you for very much! $\endgroup$ Nov 14 '13 at 18:55

You are not looking for clustering.

Instead, you are looking for frequent itemset mining. There are dozens of algorithms for this, the most widely known probably are APRIORI, FP-Growth and Eclat.


I use a simple Excel spreadsheet and load all the drawings and then do a data sort. When you do a data sort you can learn quickly which numbers will never and have never played together. I've also written Macros that can run pairs and 3 pairs.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.