Is the content about collaborative filtering in week2 ultimately Matrix factorization?

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

Hello everyone
I just finished collaborative filtering in week 2 and wondered if the formula to find optimal X, W, b that Prof. Andrew presented is ultimately matrix factorization? He just presented it in simple form??

2

Hello @anon76241992

No, it is not. The approach presented in the lecture can handle cases where r(i, j) = 0

Raymond

3

Thanks for your replying, but I don’t quite get it. Could please elaborate more, please?

4

The R matrix used in this assignment is only there to ignore entries in the Y matrix for which the user has not provided a rating.

2 Likes
5

Let’s say there are 3 movies (M1, M2, M3), and 2 users (U1, U2).

There are 2 ratings:
U1 gave M1 a rating of 0
U2 gave M3 a rating of 1

Can we form a matrix from this dataset and then do matrix factorization?

Raymond

1 Like
6

Okay, I get it now. Thanks a lot.

7

Another question. Then assumed that we have a full entries matrix with the binary labels. Do you know any implementation that instead uses matrix factorization to factorize this matrix? Will it be similar to the video “Binary labels: favs, likes and clicks”

I am highly interested in using this idea and adjusting the loss function for matrix factorization to do dimensionality reduction for binary matrix. Most MF codes I found so far only implement MSE loss which I do not think will suit the problem.

sorry for my poor english.

1 Like
8

Hello @anon76241992

If there is no missing entry, then we certainly can use a matrix factorization algorithm, as long as the computer memory permits. However, I do not have any example of matrix-factorizing a binary matrix. I have not dealt with a problem that required me to use matrix factorization.

Cheers,
Raymond

9

Then If there are no missing entries in a matrix, Prof. Andrew’s method and MF are the same whether entries are rating (inter numbers) or binary (0, 1) ??

10

If you meant “getting the same results”, then it might not be, because the results depend on the algorithms. For example, some matrix factorization gives us exact solution, but we can only hope a gradient-descent-based approach to converge. At the end of the day, we need to experiment them ourselves.

1 Like
11

Yes, thanks a lot for your reply.