Hello,
Just wrapped up week 2 of the course and enjoyed the discussion about content based filtering. I was hoping that Prof Ng would talk about the recommended dimensionality of the output of user and item neural networks. Are there any rules of thumb/methods to map the dimension of the input feature space to that of the output of said networks?
Hello @rkakade87 ,
Glad to hear you’re enjoying the course!
There isn’t a strict rule, but a common approach is to start with an embedding dimension that is smaller than the input feature space but large enough to capture the complexity of the data. A typical range might be between 50 to 300 dimensions, depending on your dataset (use cross-validation to experiment with different dimensions to find the best value). Also, computational resources must be considered, too!
Hope this helps! feel free to ask if you need further assistance!