Recommending from a large catalog

Course Q&A Machine Learning Specialization Unsupervised Learning, Recommenders, Reinforcement
1
  1. in step 1, is top 10 similar movies are selected for each of the 10 movies? So total 100 is returned?
  2. in step 1, every movie in the catalog (for example total 1000 movies) will be pre-computed for its Vm (movie feature vector). Once a new user comes in, we will take user 10 last watched movie and compute its distance to all the 999 movies in the catalog and then we rank the result. Lastly, we take top 100 from the rank and remove the duplicates. Is this description correct?
  3. in step 2 and 3, we just do a simple rank within genre and country assuming genre and country are known available features?
  4. What if it’s a new user, how do we know the 10 last watched movie?

image
image1358Ă—691 135 KB

2

Hello, @flyunicorn,

First of all, the (1), (2) and (3) there are not steps. We don’t do one after another. They are three strategies that we can apply when possible. For example, if we know the genre of all movies, then we can consider to apply strategy (2).

Within each strategy, there are two steps - retrieval and ranking. For example, strategy (3) retrieves all movies in the same country, then sort them by their popularity.

In the context of the strategies on the slide, in my understanding, for your Q1, yes and some of them may be duplicate. For Q2, mostly yes except that we have ten top-10 list instead of one top-100 list, but it doesn’t mean that your new one-top-100-list strategy must be worse but it could be less diverse. For Q3, given that they are known, instead of assuming, because if we don’t know the genres, we can’t use that strategy. For Q4, if the user just registered, then there is just no watched movie and we might want to consider other strategies.

Cheers,
Raymond

1 Like
3

Hi, please could you help me understand how exactly strategy (2) would be implemented?

Would we find the most similar movies to every movie in the user’s most-watched genres and then select top 10 per watched genre?

Or would we compute similarity between ever watched movie and candidate movie like in strategy 1) and filter 10 most similar movies that are from the same genre as user’s most watched 3 genres?