Probabilistic Principal Component Analysis
read more
Citations
Principal Component Analysis
Representation Learning: A Review and New Perspectives
Pattern Recognition and Machine Learning
Probabilistic Matrix Factorization
A survey of collaborative filtering techniques
References
The Nature of Statistical Learning Theory
Support-Vector Networks
Statistical Analysis with Missing Data
Principal Component Analysis
A training algorithm for optimal margin classifiers
Related Papers (5)
Frequently Asked Questions (8)
Q2. What is the simplest way to reduce the likelihood of a given vector?
Then consider a perturbation to this solution of the form W cW VR where is an arbitrarily small constant and the d q matrix V is given byV uiIt will be su cient to only consider those ui that are not in Uq A solution with a repeated eigenvector implies one lj becoming zero and thus a decrease in the likelihood Arbitrary permu tations of the columns of V with all valid ui thus implies that the resulting vectors vec VR are a complete orthogonal basis for the directions of interest on the likelihood surface
Q3. What is the simplest way to determine the log likelihood?
Uq may contain any of the eigenvectors of S so to identify those which maximise the likelihood the expression forW in is substituted into the log likelihood function to giveL N d ln q X j ln jdX j q j d q ln qwhere q is the number of non zero lj Di erentiating the log likelihood with respect to and substituting forW from givesd q dX j q jand soL N q X j ln j d q ln d q dX j q j A d ln dNote that implies that if rank S q as stated earlier
Q4. What is the likelihood of the data?
WTt of the data the least squares reconstruction errorE recN NX n k tn WW Ttn kis minimised when the columns ofW span the principal subspace of the data covariance matrix
Q5. What is the weight matrix for lj?
First the authors express the weight matrixW in terms of its singular value decompositionW ULVTwhere U is a d q matrix of orthonormal column vectors L diag l l lq is the q q diagonal matrix of singular values and V is a q q orthogonal matrix NowC W The authorWWT WW The authorWTWULVT The authorVLUTULVTULVTV The authorLUTUL VTUL The authorL VTThen at the stationary pointsSC W WSUL The authorL VT ULVTSUL U The authorL LFor lj equation implies that if U u u uq then each column vector uj must be an eigenvector of S with corresponding eigenvalue j such that l j j and solj jFor lj uj is arbitrary
Q6. What is the way to minimise a discarded eigenvalue?
By reference to equation the authors can deduce from this that the smallest eigenvalue must be discarded and included in the right hand term of Given the requirement that the discarded eigenvalues must be contiguous A must then be minimised when the smallest d q eigenvalues are present in the right hand term of and so L is maximised when j j f qg are the largest eigenvalues of SIt should also be noted that A is minimised with respect to q when there are fewest terms in the sum in which occurs when q q and therefore no lj is zero Furthermore L is minimised whenW which may be seen to be equivalent to the case of qIf stationary points represented by minor eigenvector solutions are stable maxima then local maximisation via an EM algorithm for example is not guaranteed to converge on the optimal solution comprising the principal eigenvectors
Q7. what is the q rank of the weight matrix?
In the context of standard PCA such a result is only attainable if q rank S For probabilistic PCA it is necessary to consider the case in which the d q smallest eigenvalues of S are identical or trivially q d because C S is attainable with min the smallest eigenvalue of S As discussed in Section W is then identi able sinceI WWT SWWT S Iwhich has a known solution at W U The authorR where U is a square matrix whose columns are the eigenvectors of S with the corresponding diagonal matrix of eigenvalues and R is an arbitrary orthogonal rotation matrixSC W W withW and C SThe authors are interested in case where C S and the model is approximate
Q8. What is the smallest eigenvalue in the spectrum of a discarded e?
Then substituting for cW gives The authorcWTcW RTKqR such thatCG SVR RTK q R VR soG C V iK q The authorRwhereiiwith i in the corresponding position to ui in V Thenvec VR Tvec G tr GTVRtr RT iK q The authorV TC VRi ki u T i C uiwhere ki is the value in Kq in the corresponding position to i Since C is positive de nite clearly uTi C ui is always positive