Hi everyone, I am trying to come up with an explanation for myself to understand/ remember the logic of recommender systems. Can the following considered a good description?
In collaborative filtering, we train the parameters ( separate vectors for a specific user and specific item) and train those parameters (or their dot product) to represent the ratings.
In content-based filtering we train the parameters of different neural network that gives us a final vector each for a specific person and item who dot product then gives us a rating.
Just to go a bit further, for those who like to focus on how the optimization problem was set up, though it sounds more complicated, a pretty equivalent way of saying your view is that
we so formulated and optimized the problem that
the errors between the outcomes of the dot product (plus bias, for the case of collaborative filtering) and the observed ratings are minimized, so that
the likelihood of the observed ratings given that trained model (consisting of the architecture including the dot product, and all of the trained parameters in that architecture) is maximized.
Either of 2 or 3 tells us what Sudhan has told us - the dot product outcomes are trained to be like the ratings.
In fact, if these are still questionable (indeed they can still be questionable to some people), depending on how it is questioned, it will then be justified to go more further.