Some Sparse Methods for High Dimensional Data
Gilbert Saporta CEDRIC,
Conservatoire National des Arts et Métiers, Paris gilbert.saporta@cnam.fr
• A joint work with
– Anne Bernard (QFAB Bioinformatics) – Stéphanie Bougeard (ANSES)
– Ndeye Niang (CNAM) – Cristian Preda (Lille)
Outline
1. Introduction
2. Sparse regression 3. Sparse PCA
4. Group sparse PCA and sparse MCA
5. Clusterwise sparse PLS regression
6. Conclusion and perspectives
1.Introduction
• High dimensional data: p>>n
– Gene expression data – Chemometrics
– etc.
• Several solutions for regression problems with all variables (PLS, ridge, PCR) ; but
interpretation is difficult
• Sparse methods: provide combinations of few
variables
• This talk:
– a survey of sparse methods for supervised
(regression) and unsupervised (PCA) problems – New propositions for large n and large p when
unobserved heterogeneity among observations occurs
• Clusterwise sparse PLS regression
2. Sparse regression
• Keeping all predictors is a drawback for high dimensional data: combinations of too many variables cannot be interpreted
• Sparse methods simultaneously shrink
coefficients and select variables, hence better
predictions
2.1 Lasso and elastic-net
The Lasso (Tibshirani, 1996) is a shrinkage and
selection method. It minimizes the usual sum of squared errors, with a bound on the sum of the absolute values of the coefficients (L1 penalty).
2
1
ˆ arg min
p
lasso j
j
β
β y Xβ
2
1
min with
p
j j
c
y - Xβ• Looks similar to ridge regression (L
2penalty)
• Lasso continuously shrinks the coefficients towards zero when c decreases
• Convex optimisation; no explicit solution
• If Lasso identical to OLS
2 2
2 2
min with
ˆridge arg min -
c
y - Xβ β
β y - Xβ β
1 p
jols j
c b
• Constraints and log-priors
– Like ridge regression, the Lasso is a bayesian regression but with an exponential prior
– j is proportional to the log-prior
( ) exp
j 2 j
f
• Finding the optimal parameter
– Cross validation if optimal prediction is needed – BIC when the sparsity is the main concern
a good unbiased estimate of df is the number of nonzero coefficients . (Zou et al., 2007)
2
2
ˆ( ) log( ) ˆ
arg min ( )
opt
y X n
n n df
• A few drawbacks
– Lasso produces a sparse model but the number of selected variables cannot exceed the number of units
• Unfitted to DNA Arrays : n(arrays)<<p(genes)
– Select only one variable among a set of correlated predictors
Elastic net
• Combines ridge and lasso penalties to select more predictors than the number of
observations (Zou & Hastie, 2005)
– L1 part leads to a sparse model
– L2 part removes the limitation on the number of retained predictors and favourises group choice
2 2 2 1 1
ˆ
en arg min
β
β y Xβ β β
2.2 Group-lasso
• X matrix divided into J sub-matrices Xj of pj
variables
• Group Lasso: extension of Lasso for selecting groups of variables (Yuan & Lin, 2007):
2
1 1
ˆ arg min
J J
GL j j j j
j j
p
β y X β β
If pj=1 for all j, group Lasso = Lasso
• Drawback: no sparsity within groups
• A solution: sparse group lasso
(Simon et al. , 2012)– Two tuning parameters: grid search
2
1 2
1 1 1 1
min
pj
J J J
j j j ij
j j j i
β y X β β
2.3. Sparse PLS
Several solutions:
Le Cao et al. (2008)
Chun & Keles (2010)
3.Sparse PCA
• In PCA, each PC is a linear combination of all the original variables : difficult to interpret the results
• Challenge of SPCA: obtain components easily interpretable (many zero loadings in principal factors)
• Principle of SPCA: modify PCA imposing lasso/elastic-net constraints to construct modified PCs with sparse loadings
• Warning: Sparse PCA does not provide a global selection of variables but a selection dimension by dimension : different from the regression
context (Lasso, Elastic Net, …)
3.1 First attempts:
• Simple PCA
– by Vines (2000) : integer loadings
– Rousson, V. and Gasser, T. (2004) : loadings (+ , 0, -)
• SCoTLASS (Simplified Component Technique – Lasso) by Jolliffe & al. (2003) : extra L
1constraints
2
1
max with 1 and
p
j j
u t
u'Vu u u'u
SCotLass properties:
• Non convex problem
usual PCA t 1 no solution
1 only one nonzero coefficient 1
t p
t
t p
Let the SVD of X be with the principal components Ridge regression:
Loadings can be recovered by regressing (ridge regression) PCs on the p variables
PCA can be written as a regression-type optimization problem
2 2
β
ˆridge arg min
β Z - Xβ β
' '
2 2
ˆ ii
ii
with
d
d
2 '
' -1 '
i,ridge i i
X X = VD V V V = I
β = X X + λ I X Xv v
+ v vi
X = UDV' Z = UD
3.2 S-PCA R-SVD by Zou et al (2006)
is an approximation to , and the ith approximated component
Produces sparse loadings with zero coefficients to facilitate interpretation
Alternated algorithm between elastic net and SVD
Sparse PCA add a new penalty to produce sparse loadings:
ˆ arg min
2 2
β Z - Xβ + β
ˆ ˆ
ˆ
i
v = β
β Xvˆ i
Lasso penalty
vi
1 1
+ β
• Loss of orthogonality
– SCotLass: orthogonal loadings but correlated components
– S-PCA: neither loadings, nor components are orthogonal
– Necessity of adjusting the % of explained variance
4. Group sparse PCA and sparse MCA
(Bernard, Guinot, S., 2012)
• Industrial context and motivation:
– Funded by R&D department of Chanel cosmetic company
– Relate gene expression data to skin aging measures
– n=500, p= 800 000 SNP’s, 15 000 genes
Data matrix X divided into J groups Xj of pj variables
Group Sparse PCA: compromise between SPCA and group Lasso
Goal: select groups of continuous variables (zero coefficients to entire blocks of variables)
Principle: replace the penalty function in the SPCA algorithm
by that defined in the group Lasso
2 2
1 1
β
ˆ arg min
β Z - Xβ β β
2
ˆ arg min
J J
GL j j pj j
β Z X β β
4.1 Group Sparse PCA
In MCA:
Selection of 1 column in the original table (categorical variable Xj )
=
Selection of a block of pj indicator variables in the complete disjunctive table
Challenge of Sparse MCA : select categorical variables, not categories
Principle: a straightforward extension of Group Sparse PCA for groups of indicator variables, with the chi-square metric . Uses s-PCA r-SVD algorithm.
XJ
1 pJ
. . . 3
XJ1 … XJpj
1 0 . . . 0
0 1 . . . 0
Original table Complete disjunctive
table
4.2 Sparse MCA
Single Nucleotide Polymorphisms
Data:
n=502 individuals
p=537 SNPs (among more than
800 000 of the original data base, 15000 genes)
q=1554 (total number of columns) X : 502 x 537matrix of qualitative variables
K : 502 x 1554 complete disjunctive table K=(K1, …, K1554)
1 block
=
1 SNP = 1 Kj matrix
Application on genetic data
Single Nucleotide Polymorphisms
Application on genetic data
Single Nucleotide Polymorphisms
Application on genetic data
λ= 0:005: CPEV= 0:32% and 174 columns selected on Comp 1
Comparison of the loadings
Application on genetic data
..
Single Nucleotide Polymorphisms
Application on genetic data
Comparison between MCA and Sparse MCA on the first plan
Properties MCA Sparse MCA
Uncorrelated Components TRUE FALSE
Orthogonal loadings TRUE FALSE
Barycentric property TRUE Partly true
% of inertia
Total inertia
j 100 tot
2 j.1,..., j-1
Z
2 1
k
j
Zj.1,..., j-11
1 1
p
j j
p
p are the residuals after adjusting for (regression projection)
Z Z Z
5. Clusterwise sparse PLS regression
• Context
– High dimensional explanatory data X (with p>>n), – One or several variables Y to explain,
– Unknown clusters of the n observations.
• Twofold aim
– Clustering: get the optimal clustering of observations, – Sparse regression: compute the cluster sparse
regression coefficients within each cluster to improve the Y prediction.
Improve high-dimensional data interpretation.
• Clusterwise (typological) regression
– Diday (1974), Charles (1977), Späth (1979)
– Optimizes simultaneously the partition and the local models
– p >nk our proposal : local sparse PLS regressions
21 1
min ( ) ( ' )
n K
k i k k i
i k
i y
1 β x
• Criterion (for a single dependent variable and G clusters)
• w loading weight of PLS component t=Xw
• v loading weight u=Yv
• c regression coefficients of Y with t: c=(T’T)-1T’y
(cluster) sparse components (cluster) residuals
' ' 2 ' 2
1
arg min with 1
G
g g g g g g g g g g g
g
X y w v w y X w c wA simulation study
• n=500, p= 100
• X vector normally distributed with mean 0 and
covariance matrix S with elements sij = cov(Xi, Xj) = ρ|i-j| , ρ ∈[0,1]
• The response variable (Y) is defined as follows:
there exists a latent categorical variable G, G∈{1,…, K}, such that, conditionally to G=i and X=x, we have :
2 /G i, ( 0i 1 1i ... pi p, i ) Y X=x N
x
x
• Signal/noise
• the n subjects are equiprobably distributed into the K clusters
• For K=2 we have chosen
2
( / ) 0.1
i
Var Y G i
• A wrong number of groups
Application to real data
• Data features (liver toxicity from mixOmics R package) – X: 3116 genes (mRNA extracted from liver)
– y: clinical chemistry marker for liver injury (serum enzymes level)
– Observations: 64 rats
• Each rat is subject to different more or less toxic acetaminophen doses (50, 150, 1500, 2000)
• Necropsy is performed at 6, 18, 24 and 48 hours after exposure (i.e., time when the mRNA is extracted from liver)
• Data pre-processing
– y and X are centered and scaled
• Twofold aim
– Improve the sparse PLS link between genes (X) and liver injury marker (y)…
– … while seeking G unknown clusters of the 64 rats.
• Batch clusterwise algorithm
– For several random allocations of the n observations into G clusters
• Compute G sparse PLS within each cluster (select and dimension through 10-fold CV)
• Assign the observation to the cluster with the minimum criterion value.
– Iterate until convergence
– Select the best initialization (i.e., which minimizes the criterion) and get:
• The allocation of the n observations into G clusters,
• The sparse regression coefficients within each cluster.
Clusterwise sparse PLS: performance vs standard spls
No cluster
• N=64
• R2=0.968
• RMSE=0.313
• =0.9 (sparse parameter)
• H= 3 dimensions
G=3 clusters of rats
• N=(12; 42; 10)
• R2=(0.99; 0.928; 0.995)
• RMSE=(0.0272; 0.0111; 0.0186)
• =( 0.9; 0.8; 0.9)
• H=(3; 3; 3)
Clusterwise sparse PLS: regression coefficients
Interpretation
• Links between genes (X) and marker for liver injury (y) are specific to each cluster
– Cluster 1: 12 selected genes – Cluster 2: 86 selected genes – Cluster 3: 15 selected genes – (No cluster: 31 selected genes)
• The cluster structure is
significantly associated with the acetaminophen dose (pvalue- Chi2Fisher= 0.0057)
– Cluster 1: all the doses
– Cluster 2: mainly doses 50, 150 – Cluster 3: mainly doses 1500, 2000
6.Conclusions and perspectives
• Sparse techniques provide elegant and efficient solutions to problems posed by high-dimensional
data. A new generation of data analysis methods with few restrictive hypothesis
• Sparse PCA and group-lasso extended towards
sparse MCA. Clusterwise extensions for large p and large n
• Research in progress:
– Criteria for choosing the tuning parameter λ
– Extension of Sparse MCA : compromise between the Sparse MCA and the sparse group lasso
developed by Simon et al. (2002)
• select groups and predictors within a group, in order to produce sparsity at both levels
– Extension to the case where variables have a block structure
– Writing a R package
References
• Bernard, A. , Guinot, C., Saporta, G. . (2012) Sparse principal
component analysis for multiblock data and its extension to sparse multiple correspondence analysis , Compstat 2012, pp.99-106,
Limassol, Chypre
• Charles, C. (1977): Régression Typologique et Reconnaissance des Formes. Thèse de doctorat, Université Paris IX.
• Chun, H. and Keles, S. (2010), Sparse partial least squares for
simultaneous dimension reduction and variable selection, Journal of the Royal Statistical Society - Series B, Vol. 72, pp. 3–25.
• Diday, E. (1974): Introduction à l’analyse factorielle typologique, Revue de Statistique Appliquée, 22, 4, pp.29-38
• Jolliffe, I.T. , Trendafilov, N.T. and Uddin, M. (2003) A modified principal component technique based on the LASSO. Journal of Computational and Graphical Statistics, 12, 531–547
• Rousson, V. , Gasser, T. (2004), Simple component analysis. Journal of the Royal Statistical Society: Series C (Applied Statistics), 53,539- 555
• Shen, H. and Huang, J. Z. (2008). Sparse principal component
analysis via regularized low rank matrix approximation. Journal of
• Simon, N., Friedman, J., Hastie, T., and Tibshirani, R. (2012) A Sparse-Group Lasso. Journal of Computational and Graphical Statistics,
• Späth, H. (1979): Clusterwise linear regression, Computing, 22, pp.367-373
• Tibshirani, R. (1996) Regression shrinkage and selection via the Lasso. Journal of the Royal Statistical Society, Series B, 58, 267-288
• Vines, S.K., (2000) Simple principal components, Journal of the Royal Statistical Society: Series C (Applied Statistics), 49, 441-451
• Yuan, M., Lin, Y. (2007) Model selection and estimation in
regression with grouped variables. Journal of the Royal Statistical Society, Series B, 68, 49-67
• Zou, H., Hastie , T. (2005) Regularization and variable selection via the elastic net. Journal of Computational and Graphical Statistics, 67, 301-320,
• Zou, H., Hastie, T. and Tibshirani, R. (2006) Sparse Principal Component Analysis. Journal of Computational and Graphical Statistics, 15, 265-286.
• H. Zou, T. Hastie, R. Tibshirani, (2007), On the “degrees of freedom”
of the lasso, The Annals of Statistics, 35, 5, 2173–2192.