On Mon, Mar 24, 2014 at 4:46 PM, Si Chen <[email protected]>wrote:
> Thanks everybody for your feedback! I thought more about it, and basically > our issue is that we have a lot of SKU's per brand, so there's not a lot of > repeat sales of the same SKU's to make SKU to SKU market basket analysis > useful. We can move it up to the brand level, but then I'm concerned that > the results would be statistically significant but not very useful. Some > brands have 5K to 10K SKU's, so a cross-sell of brand A to brand B would > still leave thousands of cross-sell candidates being suggested. > This sounds right. > > Thinking about it further, maybe what we should try to do is to divide our > catalog into small units, finer than a brand but more aggregated than an > individual SKU. Maybe if we could partition our catalog into sub-units of > 100 or so products each, then try to do market basket analysis at the > levels of these units? The products in each sub-unit would be related to > each other, but they wouldn't be necessarily be bought together. > Good idea. > Should we try to do it this way? Is there a way to partition a catalog > analytically, or would we have to do it based on human knowledge? > Well, yes and no. The only way to analytically group SKU's into reasonably sensible groups is to look at something like cooccurrence. You said that doesn't work so you have a dead end. Do you have any way to connect market baskets together? Other than that, external information will be required. That said, I wouldn't limit the analysis to a single level. If you have Brand > X > SKU, then I would consider each occurrence of a SKU to be an occurrence of X and of the Brand as well.
