0

I have very long history of user behaviour, when they choose to buy one of the 50 products. I want to take in account that if a user bought product1 two years ago and product2 yesterday, second purchase would be more relevant to the recommended items, than first one. What is the proper way to do so?

RInatM
  • 173
  • 6
  • It depends on the model/algorithm for your recommender. Question needs more detail. In any case, you can incorporate some time discounting for the product weights/scores. Choosing a suitable discount (e.g. linear, or exponential in time etc) might be a tougher choice. Hopefully you can estimate it from past data. – P.Windridge Feb 17 '15 at 16:06
  • now I use IBCF with cosine distance. How I can use weighting in this situation? – RInatM Feb 17 '15 at 16:12

1 Answers1

1

Dynamic recommender systems is a right term to google for. In the discussion above someone recommended the paper 'Collaborative filtering with temporal dynamics' by Yehuda Koren , which helped me a lot to understand the subject.

Community
  • 1
RInatM
  • 173
  • 6
  • I will update the answer as soon as I figure out the simple way to implement discussed algorithm. And still welcome other suggestions – RInatM Feb 19 '15 at 14:06