3. Outline
Expression power of matrix
Various matrix factorization methods
Application of matrix factorization
4. Curse of Dimensinality
In machine learning and others,
Dimension is a curse i.e., as they are
intensive computation, and if
dimension is high, this will worsen the
computation, that’s why we have
5. Different matrix factorization
methods
LU decomposition
Singular Value Decomposition(SVD)
Probabilistic Matrix
Factorization(PMF)
Non-negative Matrix
Factorization(NMF)
6. Application of matrix factorization
LU decomposition
◦ Solving system of equations
SVD decomposition
◦ Low rank matrix approximation
◦ Pseudo-inverse
7. Application of matrix factorization
PMF
◦ Recommendation system
NMF
◦ Learning the parts of objects
8. PMF
Consider a typical recommendation
problem
◦ Given a n by m matrix R with some
entries unknown
n rows represent n users
m columns represent m movies
Entry represent the ith user’s rating on the jth
movie
◦ We are interested in the unknown entries’
possible values
i.e. Predict users’ ratings
ij
R
9. PMF
We can model the problem as R=U’V
◦ U (k by n) is the latent feature matrix for
users
How much the user likes action movie?
How much the user likes comedy movie?
◦ V (k by m) is the latent feature matrix for
movies
To what extent is the movie an action movie?
To what extent is the movie a comedy movie?
10. PMF
If we can learn U and V from existing
ratings, then we can compute
unknown entries by multiplying these
two matrices.
Let’s consider a probabilistic
approach.
12. PMF
We want to maximize
Equivalent to minimizing
Can be solved using steepest descent
method
13. Extension to PMF
We can augment the model as long as
we have additional data matrix that
share comment latent feature matrix
14. NMF
Consider the following problem
◦ M = 2429 facial images
◦ Each image of size n = 19 by 19 = 361
◦ Matrix V = n by m is the original dataset
◦ We want to approximate V by two lower
rank matrix W (n by 49) and H (49 by m)
V ~ WH
Constraints
All entries of W and H are non-negative
15. NMF
How well can W and H approximate V
How can we interpret the result
17. Criticize of NMF
NMF doesn’t always give
parts based result
Sparseness constraints
For more information, refer
to “Non-negative matrix factorization with sparseness
constrains”