Hello,

I am using the FCHM Bond algorithm and find the bond measure to be extremely useful. I'm using this bond to identify products that cross-sell well together.

One issue I am running into is that this bond is very difficult to translate into expected volumes. For instance consider the case when a customer buys between 4 and 5 boxes of tomatoes for every 1 box of Pizza Dough. Let's say that I have a bond of 0.5 for tomatoes and dough. This would imply that for every transaction containing either tomatoes or dough id expect 50% of those transactions to include the other. I'd like to translate this into an expected number of cases of the other product sold but short of calculating an expected volume of every combination, I'm not sure how to go about solving this problem.Is there an algorithm available for this purpose? Any insight would be appreciated. Thank you!

Hi,

Thanks for using SPMF. I took a bit of time to answer because I was abroad.

Yes, the bond is a simple measure but the concept is quite interesting to find items that appear often together.

About translating patterns into volumes of sales, I do not know any algorithms that would specifically do this and I have never thought about this idea.

If we have an itemset {a1, a2 ... an} the FCHM bond algorithm can know:
- the support of a1
- the support of a2
- ...
- the support of an
- the support of {a1, a2 ... an}
- the bond of {a1, a2 ... an}

Using such information, various measures can be derived such as the bond. But how to calculate volumes? I don't know.

If we see this as a task of prediction, then maybe that some prediction models could be trained using the patterns?

Best regards,

Philippe thank you so much for your response, I appreciate it. I think there may be a potential to compute what I am looking for based on your response. We know that the

Total_Utility = Utility_a + Utility_b +... Utility_n

If we treat utility as case volume instead of dollars, and keep track of unit utilities (Utility_a through Utility_n) or unit fractions and support it would be simple to compute the decomposed mean case volumes in supporting examples. Unfortunately the unit utilities were not listed in your list of possible outputs from the algorithm. I'll dig in the code to see if I can create what I want, I just wanted to keep track of my idea in case anyone else finds it useful.

Thanks again!

-Lee

I see. Happy that you may have found a way to do what you wanted to do!

Best,

Philippe