Cost function for features and paramters for collaborative filtering?

Course Q&A Machine Learning Specialization Unsupervised Learning, Recommenders, Reinforcement
1

how does that make sense that you dont have the features and you dont have the parameters so you make a function without anything and try ML to improve that? dont we require atleast a feature like in his prevoius example; once he said he had the features, in the other he said he had the parameters now he doesnt have anything how can ML be applied to it

2

Hello @Saim_Rehman
In Collaborative Filtering all predictions are based on data available. From Data only all the values of parameters and features are estimated to filter out basically the content of similar type. Image of Collaborative Filtering
Hope this helps
Please do ask if you still have any queries.
Riya

3

yes that is my point, andrew says that if we donot have the features and the w too use the combined formula, how can we then make the model learn if we donot have an input?

4

Hello @Saim_Rehman,

Given what you have shared with us in this thread, I recommend you to read this whole thread first, and then let us know if you have any specific questions.

Btw, we are providing the ratings to the model at training :wink:

Raymond

5

yes we just have the predicted ratings, while in course 1 or some other algorithm, we had atleast an input x and an output y, but here there is no input no w and no x so how does the algo work

6

Before continued, how does linear regression work in course 1: Did it bring w, by gradient descent, to a set of values such that wx +b mostly fits to y?

7

yes?
there was an input x everytime but this time he says what if we dont have the features

8

Ok! Then this time, gradient descent also brings x (together with w) such that wx +b mostly fits to y.

If x is given, good, then gradient descent has less work to do, otherwise, it has more. You need to give up the idea that x must be provided.

Note the following difference -

  1. In collaborative filtering, a x (w as well) is the quality of an entity (user/movie)

  2. In those linear regression examples for course one, a x is a sample of an environment, and only w describes the quality of the environment (the “entity”).

We model entities (environment).

The above difference may takes a millisecond, or sometimes, up to years to take in. It just requires the right time to come. :wink:

Raymond

1 Like
9

There will be times that y is completely not given, only you won’t see that in the MLS. In some real world problem, we need to do Monte Carlo simulation to generate that y ourselves. I just hope that you will leave that flexibility for different things. Data Science is an amazing world that attempts to deal with all kinds of problem (including cases when x and y are given, only y is given, only x is given, …), so be ready for that, give yourself time to take things in, and I feel excited for you :wink: :raised_hands:

Cheers,
Raymond