1
$\begingroup$

I am currently working on a best way to represent product interest score based on page views.

Working on an ecommerce use case having homepage, search page, and product page interaction data available. I want to build a customer interest score on a product and work on an effective way to represent it.

Simplest metric would be summing up all the interaction associated with the product i.e. home + search + product page i.e. if a product A was seen by 5 users on Home, 10 users in search and 2 users in product page, the total count would be 17. However, my theory is that we need to give more bias/weight to the product page since customers have specifically viewed this page as compared all the other product published in home and search page. Something like:

(5x + 10y + 2z)/(x+y+z), where x,y,z are the weights associated with home, search, and item page respectively. In my theory, z > y > x

How to best represent this metric by weighting the pages accordingly?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
0
$\begingroup$

If the product seen on home or search page can be just a result of randomly scanning the page by the user and might not necessarily mean that the user expresses an interest for the product, while something that has been seen on the product page means that the user clicks into that page, I think it would be interesting to represent this information in one-dimension by applying PCA, and check if the most variance in the component does come from the variable z. If so, you could probably use the PCA component scores as an indication of interest level.

Improve this answer
$\endgroup$
3
  • $\begingroup$ Thank you! The catch is that users can add to cart from any of the mentioned pages. It is likely that users will visit the product page before buying anything, but they can make a purchase even without visiting the PDP $\endgroup$
    – Cooper
    Commented Dec 21, 2021 at 6:00
  • $\begingroup$ do you have access to the labels, whether the customer eventually bought the product? why not use this information as well? $\endgroup$
    – user344849
    Commented Dec 21, 2021 at 13:06
  • 1
    $\begingroup$ I do, but that would eventually bias toward the items that customer actually bought, which is relatively a low number as compared to all the items in catalog. On the flipside, we want to know why didn't the customer buy despite having so many views (an aggregated metric) -- is it the price? is it the product picture? shipping promise? etc. For this reason, I want to get a normalized views metric $\endgroup$
    – Cooper
    Commented Dec 21, 2021 at 16:26

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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