Goal: Generate a list of 100 products per vertical (e.g. fashion, electronics) that the teams should source, discount, and list on the website over a specific period. You may assume all customers are online only.

My thinking so far:

  1. Predict the customers that will come to the website during the specific period (time series). Only 30-40% customers return YOY.
  2. Understand what they want (use search data, add to basket but didn't checkout etc). Potentially segment it further by looking at those customers who generate the highest revenue in general vs one off purchasers.
  3. Further filter those products that these customers add (or take from 'viewed or saved' state to the 'checkout' state) once the product is on deal.
  4. Potentially use clustering to recommend products similar to those from step 3 but that have never been on a deal?

I cannot influence the amount of discount. I can only influence what products we source. Therefore I want to source the ones with the highest potential of purchasers from the customer.

Any thoughts on the approach above?


Clustering is not well suited for prediction.

Use a recommender system instead.

  • 1
    $\begingroup$ Can you kindly link me to an example (papers / links / blogs etc with an example would be awesome). Sounds like collaborative filtering is needed? en.wikipedia.org/wiki/Recommender_system $\endgroup$ – lseactuary Mar 4 '18 at 14:49
  • $\begingroup$ I guess the complication here is what products to recommend to source on discount which is different to just 'any product'. Taking a subset of products that have been on discount before would skew the results. Customers are also changing but the recommender system would be based on all customers in say the past 12 months right (only about 15% are new customers)? $\endgroup$ – lseactuary Mar 4 '18 at 17:31
  • $\begingroup$ Sorry, I do not have link recommendations. $\endgroup$ – Has QUIT--Anony-Mousse Mar 4 '18 at 17:40
  • $\begingroup$ Have read up on this and there are some papers focused on algorithms to surface "top n items". The issue is these will surface "all" items vs items that would perform / are a "must have" on discount. That is why I can't directly apply the algorithm and therefore was thinking of the clustering to get "like" items. What do you think? $\endgroup$ – lseactuary Mar 7 '18 at 22:03
  • $\begingroup$ Use the 'on sale' subset only. $\endgroup$ – Has QUIT--Anony-Mousse Mar 8 '18 at 7:28

If you know the leaved basket products or the purchase the user did in the past, you can use Association Rule Mining to find the most probable purchase for the customers. You can select the top-n products/categories and develop the discount strategy for them. Here is good post about Basker Analysis with Assosative Rules. Unless you have very specific sales data ( like B2B with small amount of clients and low purchase volume), you should uncover some insights.


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.