2
$\begingroup$

How can you calculate better video recommendations for a SmartTV app which is used by multiple users in a household. I don't know which user is currently watching the video because the account for the app is shared. So the current user might get a recommendation based on video usage of an other user.

I was thinking of using principal component analysis to find out how many users live in the household and which user is likely to watch the video now. I could map the usage data to this specific user and then build recommendations for this hypothetical user.

What would be the best approach to solve this problem. Are there best practices or papers which describe a solution?

|improve this question
$\endgroup$
  • $\begingroup$ Your statement assumes a community-based recommender. Why not use a content-based one instead? That would provide you a quicker path to a solution $\endgroup$ – I_Play_With_Data Mar 8 '18 at 17:24
  • $\begingroup$ There is some literature on this neglected problem under the rubric of shared account recommenders. The best results will be obtained when the user reveals herself, as in Netflix today. Next best is when they reveal how many users there are but not which one, in which case you have a classification task. The last case in which you know neither you have a clustering task. A nonparametric Bayesian approach is possible. You need to learn a representation that allows you to distinguish users. So you have two problems: determine the users from the accounts, then provide them recommendations. $\endgroup$ – Emre Mar 8 '18 at 17:35
1
$\begingroup$

I was interestedin this problem awhile back ago. I still have this paper and should serve as a good primer.

Guess Who Rated Thos Movie:...

|improve this answer
$\endgroup$
0
$\begingroup$

Generally, there is a big risk to the 2-stage approach of (1) classifying who watches at the moment and (2) serving recommendations to the predicted user: if you get stage 1 wrong your recommendations will be terribly off.

I think you should consider calculating and serving recommendations to the complete household, i.e. a mixture of recommendations to the Dad, Daughter, etc. Additionally, you could change the order based on your user classification. This would give you superior personalization but you won't be completely off if you predict the wrong user.

As for specific strategies you could check early Netflix Papers and Blog Posts. If I remember correctly they used to have only one user per account.

Hope this helps!

|improve this answer
$\endgroup$

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.