Bayesian selection of best subsets in high-dimensional regression

Bayesian selection of best subsets in high-dimensional
regression
Shiqiang Jin
Department of Statistics
Kansas State University, Manhattan, KS
Joint work with
Gyuhyeong Goh
Kansas State University, Manhattan, KS
July 31, 2019
Shiqiang Jin (Department of Statistics Kansas State University, Manhattan, KS Joint work with Gyuhyeong Goh Kansas State University, Manhattan,
Bayesian selection of best subsets in high-dimensional regression July 31, 2019 1 / 25

Bayesian linear regression model in High-dimensional Data
Consider a linear regression model
y = Xβ + , (1)
where y = (y1, . . . , yn)T
is a response vector, X = (x1, . . . , xp) ∈ Rn×p
is a
model matrix, β = (β1, . . . , βp)T
is a coefficient vector and ∼ N(0, σ2
In).
We assume p n, i.e. High-dimensional data.
We assume only a few number of predictors are associated with the response,
i.e. β is sparse.

Bayesian linear regression model in High-dimensional Data
To better explain the sparsity of β, we introduce a latent index set
γ ⊂ {1, . . . , p} so that Xγ represents a sub-matrix of X containing xj , j ∈ γ.
e.g. γ = {1, 3, 4} ⇒ Xγ = (x1, x3, x4).
The full model in (1) can be reduced to
y = Xγβγ + . (2)

Priors and marginal posterior distribution
Given k, it can be shown that
π(γ|y) ∝
(τ−1
)
|γ|
2
|XT
γXγ + τ−1I|γ||
1
2 (yT
Hγy + bσ)
aσ+n
2
I(|γ| = k)
≡ g(γ)I(|γ| = k),
where Hγ = In − Xγ(XT
γXγ + τ−1
I|γ|)−1
XT
γ.
Hence, g(γ) is our model selection criterion.

Best subset selection Algorithm
Note that our goals are to obtain (i) k most important predictors out of p
k

models; (ii) a single best model from among 2p
candidate models.
Hence, this can be fallen into the description of best subset selection as
follows:
(i) Fixed size: for k = 1, 2, . . . , p, select the best subset model by
Mk = argγ max
|γ|=k
g(γ)
from p
k

candidate models.
(ii) Varying size: Pick the single best model from M1, . . . , Mp.
Challenge of best subset selection:
e.g. (i) p = 1000, k = 5, 1000
5

≈ 8 × 1012
; (ii) p = 40, 240
≈ 1012
.

Neighborhood Search
To avoid the exhaustive computation, we resort to the idea of Neighborhood
Search proposed by Madigan et al. (1995) and Hans et al. (2007).
Let γ(t)
be a current state of γ, |γ(t)
| = k is model size.
Addition neighbor: N+(γ(t)
) = {γ(t)
∪ {j} : j /
∈ γ(t)
}; model size?
Deletion neighbor: N−(γ(t)
) = {γ(t)
{j0
} : j0
∈ γ(t)
}; model size?
e.g. Suppose p = 4, k = 2. Let γ(t)
= {1, 2}, then
N+(γ(t)
) = {{1, 2, 3}, {1, 2, 4}},
N−(γ(t)
) = {{1}, {2}}.

Hybrid best subset search with a fixed k
Note our Goal (i) is to find γ̂ = argγ max|γ̂|=k g(γ).
1. Initialize γ̂ s.t. |γ̂| = k.
2. Repeat #deterministic search:local optimum
Update γ̃ ← arg maxγ∈N+(γ̂) g(γ) ; # N+(γ̂) = {γ̂ ∪ {j} : j /
∈ γ̂}
Update γ̂ ← arg maxγ∈N−(γ̃) g(γ); # N−(γ̃) = {γ̃ {j} : j ∈ γ̃}
until convergence.
In Step 2, we have p + 1 many candidate models in all γ ∈ N+(γ̂) and all
γ ∈ N−(γ̃)) in each iteration.
compute g(γ) p + 1 times in each iteration.

1st Key features of proposed algorithm
g(γ) =
(τ−1
)
|γ|
2
|XT
γXγ + τ−1I|γ||
1
2 (yT
Hγy + bσ)
aσ+n
2
,
where Hγ = In − Xγ(XT
γXγ + τ−1
I|γ|)−1
XT
γ.
compute inverse matrix and determinant p + 1 times.
We propose the following formula and we can show that evaluating all
candidate models in addition neighbor can be done simultaneously in this
single computation.
g(N+(γ̂)) =

(yT
Hγ̂y + bσ)1p −
(XT
Hγ̂y)2
τ−11p + diag(XT
Hγ̂X)
− aσ+n
2
×

τ−1
1p + diag(XT
Hγ̂X)
−1/2
, (3)
Similarly for g(N−(γ̂)).

Update γ̃ ← arg max g(N+(γ̂)) ; # N+(γ̂) = {γ̂ ∪ {j} : j /
∈ γ̂}
Update γ̂ ← arg max g(N−(γ̃)); # N−(γ̃) = {γ̃ {j} : j ∈ γ̃}
until convergence.
3. Set γ(0)
= γ̂.
4. Repeat for t = 1, . . . , T: #stochastic search:global optimum
Sample γ∗
∼ π(γ|y) = {g(γ)}
P
γ
{g(γ)}
I{γ ∈ N+(γ(t−1)
)};
Sample γ(t)
∼ π(γ|y) = {g(γ)}
P
γ
{g(γ)}
I{γ ∈ N−(γ∗
)};
If π(γ̂|y) π(γ(t)
|y), then update γ̂ = γ(t)
, break the loop, and go to 2.
5. Return γ̂.
Note that all g(γ) are computed simultaneously in their neighbor space.

2nd Key features of proposed algorithm
To avoid staying in the local optimum for a long time in step 4, we use the
escort distribution.
The idea of escort distribution (used in statistical physics and
thermodynamics) is introduced to stimulate the movement of Markov chain.
An escort distribution of g(γ) is given as
{g(γ)}α
P
γ{g(γ)}α
, α ∈ [0, 1]

Escort distributions
e.g. Consider 3 candidate models with its probability:
If α = 1,
{g(γ)}α
P
γ{g(γ)}α
=





0.02 model 1
0.90 model 2
0.08 model 3
.
1 2 3
Escort
probability
0.0
0.2
0.4
0.6
0.8
1.0
α=1
0.02
0.9
0.08
1 2 3
Escort
probability
0.0
0.2
0.4
0.6
0.8
1.0
α=0.5
0.1
0.69
0.21
1 2 3
Escort
probability
0.0
0.2
0.4
0.6
0.8
1.0
α=0.05
0.3
0.37
0.33
Figure: Escort distributions of πα(γ|y).

Update γ̃ ← arg max g(N+(γ̂)) ; # N+(γ̂) = {γ̂ ∪ {j} : j /
∈ γ̂}
Update γ̂ ← arg max g(N−(γ̃)); # N−(γ̃) = {γ̃ {j} : j ∈ γ̃}
until convergence.
3. Set γ(0)
= γ̂.
4. Repeat for t = 1, . . . , T: #stochastic search:global optimum
Sample γ∗
∼ πα(γ|y) = {g(γ)}α
P
γ
{g(γ)}α
I{γ ∈ N+(γ(t−1)
)}; # α ∈ [0, 1]
Sample γ(t)
∼ πα(γ|y) = {g(γ)}α
P
γ
{g(γ)}α
I{γ ∈ N−(γ∗
)};
If π(γ̂|y) π(γ(t)
|y), then update γ̂ = γ(t)
, break the loop, and go to 2.
5. Return γ̂.
Note that all g(γ) are computed simultaneously in their neighbor space.

Best subset selection with varying k
Note that Goal (ii): a single best model from among 2p
candidate models.
We extend ”fixed” k to varying k by assigning a prior on k.
Note that the uniform prior, k ∼ Uniform{1, . . . , K}, tends to assign larger
probability to a larger subset (see Chen and Chen (2008)).
We define
π(k) ∝ 1/

p
k

I(k ≤ K).

Hybrid best subset search with varying k
Bayesian best subset selection can be done by maximizing
π(γ, k|y) ∝ g(γ)/

p
k

(4)
over (γ, k).
Our algorithm proceeds as follows:
1. Repeat for k = 1, . . . , K:
Given k, implement the hybrid search algorithm to obtain best subset model
γ̂k .
2. Find the best model γ̂∗
obtained by
γ̂∗
= arg max
k∈{1,...,K}

g(γ̂k )/

p
k

. (5)

Consistency of model selection criterion
Theorem
Let γ∗ indicate the true model. Define Γ = {γ : |γ| ≤ K, γ 6= γ∗}. Assume that
p = O(nξ
) for ξ ≥ 1. Under the asymptotic identifiability condition of Chen and
Chen (2008), if τ → ∞ as n → ∞ but τ = o(n), then the proposed Bayesian
subset selection possesses the Bayesian model selection consistency, that is,
π(γ∗|y) max
γ∈Γ
π(γ|y) (6)
in probability as n → ∞.
As n → ∞, the maximizer of π(γ|y) is the true model based on our model
selection criterion.

Simulation study
Setup
For given n = 100, we generate the data from
yi
ind
∼ Normal
p
X
j=1
βj xij , 1
!
,
where
I (xi1, . . . , xip)T iid
∼ Normal(0p, Σ) with Σ = (Σij )p×p and Σij = ρ|i−j|
,
I βj
iid
∼ Uniform{−1, −2, 1, 2} if j ∈ γ and βj = 0 if j /
∈ γ.
I γ is an index set of size 4 randomly selected from {1, 2, . . . , p}.
I We consider four scenarios for p and ρ:
(i) p = 200, ρ = 0.1, (ii) p = 200, ρ = 0.9,
(iii) p = 1000, ρ = 0.1, (iv) p = 1000, ρ = 0.9.

Simulation study
Results (high-dimensional scenarios)
Table: 2, 000 replications; FDR (false discovery rate), TRUE% (percentage of the true
model detected), SIZE (selected model size), HAM (Hamming distance).
Scenario Method FDR (s.e.) TRUE% (s.e.) SIZE (s.e.) HAM (s.e.)
p = 200 Proposed 0.006 (0.001) 96.900 (0.388) 4.032 (0.004) 0.032 (0.004)
ρ = 0.1 SCAD 0.034 (0.002) 85.200 (0.794) 4.188 (0.011) 0.188 (0.011)
MCP 0.035 (0.002) 84.750 (0.804) 4.191 (0.011) 0.191 (0.011)
ENET 0.016 (0.001) 92.700 (0.582) 4.087 (0.007) 0.087 (0.007)
LASSO 0.020 (0.002) 91.350 (0.629) 4.109 (0.009) 0.109 (0.009)
p = 200 Proposed 0.023 (0.002) 88.750 (0.707) 3.985 (0.006) 0.203 (0.014)
ρ = 0.9 SCAD 0.059 (0.003) 74.150 (0.979) 4.107 (0.015) 0.480 (0.022)
MCP 0.137 (0.004) 55.400 (1.112) 4.264 (0.020) 1.098 (0.034)
ENET 0.501 (0.004) 0.300 (0.122) 7.716 (0.072) 5.018 (0.052)
LASSO 0.276 (0.004) 15.550 (0.811) 5.308 (0.033) 2.038 (0.034)

Simulation study
Results (ultra high-dimensional scenarios)
Table: 2, 000 replications; FDR (false discovery rate), TRUE% (percentage of the true
model detected), SIZE (selected model size), HAM (Hamming distance).
Scenario Method FDR (s.e.) TRUE% (s.e.) SIZE (s.e.) HAM (s.e.)
p = 1000 Proposed 0.004 (0.001) 98.100 (0.305) 4.020 (0.003) 0.020 (0.003)
ρ = 0.1 SCAD 0.027 (0.002) 87.900 (0.729) 4.145 (0.010) 0.145 (0.010)
MCP 0.031 (0.002) 86.550 (0.763) 4.172 (0.013) 0.172 (0.013)
ENET 0.035 (0.002) 84.850 (0.802) 4.181 (0.013) 0.206 (0.012)
LASSO 0.014 (0.001) 93.850 (0.537) 4.073 (0.007) 0.073 (0.007)
p = 1000 Proposed 0.023(0.002) 89.850 (0.675) 4.005 (0.005) 0.190 (0.013)
ρ = 0.9 SCAD 0.068 (0.003) 74.250 (0.978) 4.196 (0.014) 0.493 (0.023)
MCP 0.152 (0.004) 53.750 (1.115) 4.226 (0.017) 1.202 (0.035)
ENET 0.417 (0.005) 0.150 (0.087) 6.228 (0.068) 4.089 (0.043)
LASSO 0.265 (0.004) 19.500 (0.886) 5.139 (0.029) 1.909 (0.035)

Real data application
Data description
We apply the proposed method to Breast Invasive Carcinoma (BRCA) data
generated by The Cancer Genome Atlas (TCGA) Research Network
http://cancergenome.nih.gov.
The data set contains 17, 814 gene expression measurements (recorded on
the log scale) of 526 patients with primary solid tumor.
BRCA1 is a tumor suppressor gene and its mutations predispose women to
breast cancer (Findlay et al., 2018).

Results (based on 4, 000 genes)
Our goal here is to identify the best fitting model for estimating an
association between BRCA1 (response variable) and the other genes
(independent variables).
BRCA1 = β1 ∗ NBR2 + β2 ∗ DTL + . . . + βp ∗ VPS25 + .
Results:
Table: Model comparison
# of selected PMSE BIC EBIC
Proposed 8 0.60 984.45 1099.50
SCAD 4 0.68 1104.69 1166.47
MCP 4 0.68 1104.69 1166.47
ENET 5 0.68 1110.65 1186.25
LASSO 4 0.68 1104.69 1166.47

Results (cont.)
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.582
0.001
0.001
0.001
0.001
0.001
0.001
0.001
0.001
RPL27
C10orf76
DTL
NBR2
C17orf53
TUBG1
CRBN
YTHDC2
VPS25
CMTM5
TUBA1B
LASSO ENET MCP SCAD Proposed
Methods
Genes
−0.1
0.0
0.1
0.2
0.3
Coefficient
Figure: Except C10orf76, 7 genes are documented as diseases-related genes

Concluding remarks
Parallel computing is applicable to our algorithm with varying k.
The proposed method can be extended to multivariate linear regression
models, binary regression models, and multivariate mixed responses models
(in progress).

REFERENCES
Hans, C., A. Dobra, and M. West (2007). Shotgun stochastic search for “large p” regression.
Journal of the American Statistical Association 102(478), 507–516.
Chen, J. and Z. Chen (2008). Extended bayesian information criteria for model selection with
large model spaces. Biometrika 95(3), 759–771.
Findlay, G. M., R. M. Daza, B. Martin, M. D. Zhang, A. P. Leith, M. Gasperini, J. D. Janizek,
X. Huang, L. M. Starita, and J. Shendure (2018). Accurate classification of brca1 variants
with saturation genome editing. Nature 562(7726), 217.
Madigan, David, Jeremy York, and Denis Allard (1995). Bayesian Graphical Models for Discrete
Data. International Statistical Review / Revue Internationale De Statistique 63(2), 215-232.

Contact: jinsq@ksu.edu
THANK YOU

Bayesian selection of best subsets in high-dimensional regression

More Related Content

What's hot (20)

Similar to Bayesian selection of best subsets in high-dimensional regression (20)

Recently uploaded (20)

Bayesian selection of best subsets in high-dimensional regression