l*********s
发帖数: 5409 | 4 all right>
Below is the reference paper of the original Bayesian probability matrix
factorization method for the Netflix price data set(
http://citeseerx.ist.psu.edu/viewdoc/download?
doi=10.1.1.127.6198&rep=rep1&type=pdf )
Essentially, this is a Bayesian flavored factor analysis, with user
preference matrix as the factor, and latent movie feature matrix as the
factor loading.
My question are two fold:
1) Since the factorization is not unique due to rotation invariance, I
wonder what will be the impact on MCMC? Say, will lacking of uniqueness
results in multimodal posterior distribution/ wider marginal
distribution of
hyperparamters etc. ?
2) how to predict ratings for multiple movie-user combinations from the
simulated samples. Posterior mode seems more attractive than mean, but
estimate this vector as efficiently as possible?
ps. The article mentions L1 and L2 regularization, could someone give
a
brief description of the concept?
|
