I once modeled user journeys on a website using fancy ML models that honored sequence information, i.e., order of page visits, only to be beaten by bag-of-words (i.e., page url becomes a vector dimension, but order is lost) decision tree model, which was supposed to be my baseline.
What I had overlooked was that journeys on that particular website were fairly constrained by design, i.e., if you landed on the home page, did a bunch of stuff, put product X in the cart - there was pretty much one sequence of pages (or in the worst case, a small handful) that you'd traverse for the journey. Which means the bag-of-words (BoW) representation was more or less as expressive as the sequence model; certain pages showing up in the BoW vector corresponded to a single sequence (mostly). But the DT could learn faster with less data.
Comments
I once modeled user journeys on a website using fancy ML models that honored sequence information, i.e., order of page visits, only to be beaten by bag-of-words (i.e., page url becomes a vector dimension, but order is lost) decision tree model, which was supposed to be my baseline.
What I had overlooked was that journeys on that particular website were fairly constrained by design, i.e., if you landed on the home page, did a bunch of stuff, put product X in the cart - there was pretty much one sequence of pages (or in the worst case, a small handful) that you'd traverse for the journey. Which means the bag-of-words (BoW) representation was more or less as expressive as the sequence model; certain pages showing up in the BoW vector corresponded to a single sequence (mostly). But the DT could learn faster with less data.