|
I have a collection of videos and watch times for users. I also have a bunch of attributes of the videos and users and watch histories that I will use as predictors/features. I am interested in recommending personalized videos to a user with the goal of maximizing the watch time. How would I go about formulating the problem? I can think of two ways:
Does my approach make sense? If not, please suggest any other methods. PS: I am interested in specifically trying to maximize watch time and model this as a regression problem. I know there are other ways to recommend related items (association-rules, SVD, netflix algorithms etc.) |
|
There is no one true right answer to this question, as it really depends on your business needs. Do you want to maximize page views (because one page view is one ad impression)? Do you want to recommend shorter videos that people will watch almost all the way (option 2) or longer videos people might give up halfway through (option 1)? |
|
I think you should look at poisson regression on the rate (rather than logistic regression) Wikipedia:Poisson You are looking for something that takes into account that a rate must be between 0 and 1, ( but it doesn't have to be a probability) Could you clarify what you are concerned about? I think what you are asking is watch time as Linear regression: Pros - directly modelling the quantity you want to maximise... Cons: not taking Film length into account ie Lin regression could predict watch more than film |
