I have a problem formulating a requirement into a statistical learning problem with reportable accuracy and was hoping to tap into your expertise. The supermarket is trying to build a model that identifies customers to increase the purchase of a particular product (chocolate vs. potato wedges vs. chicken etc.), to be used for the different holidays.

Take for example Christmas is coming, and the supermarket would like to target customers who have not been buying potato wedges regularly to promote potato wedges to them. Come Easter, using the same model, the supermarket would like to target customers who have not been buying chocolate regularly to promote chocolate to them.

The initial attempt was to build an Apriori / market basket analysis model. I built a dataset that encodes regular purchase of various products per client and feed this into the Apriori model.

Modelling Data Set

Subsequently, come Christmas, I could extract the list of clients that have not been purchasing Potato Wedges regularly, and append the probability of uptake using the antecedents rules from the apriori model. For example, customers who display {chocolate = Y, chicken = Y} have a 80% probability of uptake. But that doesn’t mean that customers who display {chocolate = Y, chicken = Y, Potato Wedges = N} have an 80% of wedges uptake in the Christmas campaign right?

In this case, how would you suggest we measure the accuracy rate of the apriori model after each campaign is implemented? (Customers who bought Potato Wedges) / (All customers who are predicted to buy Potato Wedges) wouldn't work because the subset of clients is {chocolate = Y, chicken = Y, Potato Wedges = N} and not {chocolate = Y, chicken = Y}.

Would you recommend a change in approach?


Your Answer

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

Browse other questions tagged or ask your own question.