1、Abstract
PRMF:formulated with a Laplace error and a Gaussian prior which correspond to an l1loss and an l2regularizer.
2、MF mode
minU∈Rm×r,V∈Rn×r∥W⊙(Y−UVT)∥aa+λu2∥U∥22+λv2∥V∥22
3、PRMF
3.1 Probabilistic Interpretation
Probabilistic Interpretation:corresponding to a MAP estimation problem.
Yuij|λuvij|λv===UVT+EN(uij|0,λ−1u)N(vij|0,λ−1v)
suppose eij is sampled from Laplace distribution L(eij|0,λ) :
p(E|λ)=(λ2)mnexp{−λ∥E∥1}
from Bayes’, p(U,V|Y,λ,λu,λv)∝p(Y|U,V,λ)p(U|λu)p(V|λv) :
logp(U,V|Y,λ,λu,λv)=−λ∥Y−UVT∥aa+λu2∥U∥22+λv2∥V∥22
3.2 Model Reformulation
Laplace distribution:
p(z|u,α2)=α22exp(−α2|z−u|)
is equivalently expressed as a scaled mixture of Gaussians:
L(z|u,α2)=∫∞0N(z|u,τ)Expon(τ|α2)dτ
where Expon(τ|α2) is:
p(τ|α2)=α22exp(−α2τ2)
Matrix
T=[τij]∈Rm×n
, where each element
τij
is a latent variable with exponential prior for the corresponding
yij
.
yij|U,V,Tuij|λuvij|λvτij|λ∼∼∼∼N(yij|uTivj,τij)N(uij|0,λ−1u)N(vij|0,λ−1v)Expon(τij|λ/2)
3.3 Optimization
CEM: Each CEM iteration consists of two EM steps, namely, updating
V
while fixing
Q(V|θˆ)=ET[logp(V|Uˆ,Y,T)|Y,θˆ]
where T is the missing data and