Collaborative filtering feature vector for movie

Hi, at the start of collaborative filtering lecture, there is this example on movie ratings and the feature vector for movie having size 2 ie X1 (romance) and X2 (action). In real application, how do we decide the size of this “X”?
Thank you

This is a hyperparameter tuning question. Please refer to the course 2 of the Deep learning specialization for the details of the “how”, The idea is you will have a cv set which your model did not train on for the purpose of evaluation, and you pick the “size” that achieves the best evaluation score.

Thanks for the explanation. Do we ever get any insight into what the features x are, e.g. romance, action, etc. for movies. Or are the x features all unlabeled?

Hello @Daniel_Fourie,

All unlabelled. NN being a “universal approximator”, the good thing is that it can fit to any function, but it is also a bad thing that not every function carries an interpretable meaning like “movie type”. The NN chooses only functions that minimizes the cost, not functions that produce interpretable meaning, so it is quitely unlikely that we can label any of the features with a physical meaning.

Cheers,
Raymond

1 Like