0
$\begingroup$

I am attempting to predict whether an individual is likely to complete an action given a set of input data about their actions on an ecommerce website.

Some of that data could be considered "factual". For instance, their name, how much they've spent in total, products they've bought, time since they last purchased.

Some of the data however could be considered "event" based. For instance, the time between two purchases, or the amount of time passed since the last action, or the offers on sale right now, or product they've bought.

I believe that in some cases the factual data will influence the action rate (e.g. someone has previously bought a lot, suggesting loyalty to the site) and in other cases the event data will influence it (e.g. today there is a 20% off sale site wide, suggesting more users are likely to ).

Is it better to learn off the event data (where each event is considered in isolation, and you potentially lose some of the aggregated data) or the aggregate data (where you potentially lose some of the nuance between events).

UserID,Timestamp,ThisOrderValue,TotalOrderValueEver,Offers,TimesActionPerformed User1,Time0,£100,£100,20%,0 User1,Time1,£40,£140,15%,1 User1,Time2,£30,£170,15%,1 User1,Time3,£100,£270,NULL,2 User2,Time0,£10,£10,15%,0 User2,Time1,£100,£110,NULL,0

Or should I go for something that aggregates everything together, like this:

UserID,FirstSeen,LastSeen,FirstOrderValue,LastOrderValue,TotalOrderValueEver,OffersSeen,TimesPerformedAction User1,Time0,Time2,£100,£30,£270,20%;15%,2 User2,Time0,Time1,£10,£100,£110,15%,0

My assumption would be that the second is correct to learn from, but throws away possible relevant data (such as the connection between the 20% off and the first purchase, or that they didn't perform the action until their second purchase).

Which makes more sense?

Improve this question
$\endgroup$
4
  • $\begingroup$ What makes you think using the raw event data causes you to "lose some of the aggregated data"? $\endgroup$
    – Kodiologist
    Commented Aug 27, 2017 at 19:44
  • $\begingroup$ Possibly a misguided hunch! I'm worried that if each event is considered in isolation (e.g. Event1,Order1=£10,Event2,Order2=£20,etc) that I'd lose something that the aggregate includes, such as the fact that Event2 triggered the Action I want because the first order (Order1) had some property to it. $\endgroup$
    – edhgoose
    Commented Aug 27, 2017 at 19:50
  • $\begingroup$ If I understand correctly, your concern is that there could be lagged effects among the events: that is, features of one event could influence the outcome of a later event. Is that correct? Because I understand why you might have trouble seeing how to estimate a lagged effect with the event data, but not how the aggregate data would help. $\endgroup$
    – Kodiologist
    Commented Aug 27, 2017 at 19:57
  • $\begingroup$ Yes, I think that's what I mean - I believe that events affect later events. I believe the aggregate data would help because someone who has high purchase frequency or high average order volume is more likely to complete the action. And so I wonder if event data is best, enriched with aggregate data? $\endgroup$
    – edhgoose
    Commented Aug 27, 2017 at 20:07

1 Answer 1

Reset to default
0
$\begingroup$

It sounds like your best bet is to use the event data, but supplement it with lagged covariates. For example, if you want to investigate the effect of the previous order value on some characteristic of the current order, then add a column PreviousOrderValue. Or you could add a column for the total order value over all the subject's orders in the previous 30 days. There's a lot of possibilities here; books on the analysis of longitudinal data or time series are good places to get an overview.

Improve this answer
$\endgroup$

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.