FAMILY: Framework for Modeling Interactions with Strong Heredity

Description Usage Arguments Details Value References See Also Examples

View source: R/ADMM_0.1.19.R

Description

This function runs the main algorithm presented in Haris, Witten and Simon (2014) for fitting an interaction model with strong heredity.

Usage

1
2
3
4
FAMILY(X, Z, Y, lambdas , alphas, family = c("gaussian","binomial"), 
    rho = 1, B = NULL, norm = "l2", quad = TRUE,iter=500, 
    e.abs = 1e-3, e.rel = 1e-3,
    maxiter.B = 50, tol.B = 1e-04, verbose = FALSE)

Arguments

X

A n x p_1-matrix of covariates X.

Z

A n x p_2-matrix of covariates Z. The number of rows of this matrix must coincide with that of X. For most cases we have X=Z.

Y

The response vector of length n. This has to be a numeric vector. For the case of logistic regression the response variable must be a binary vector.

lambdas

The vector of different penalty parameters λ for which we wish to evaluate the function. For details see Haris, Witten and Simon (2014).

alphas

The second tuning parameter to control the magnitude of penalties on groups Of variables versus individual interaction terms. The values of this vector must be in the interval [0,1]. The output will fit the model for a grid of α and λ values.

family

A character string specifying the type of model to fit. "gaussian" for modeling continuous variables via linear regression (default), "binomial" for logistic regression.

rho

The starting value of ρ>0, the augmented Lagrangian parameter.

B

Initial (p_1+1) x (p_2+1) matrix of coefficients, B. B[1,1] is the intercept, B[1,-1] and B[-1,1] are the main effects of Z and X, respectively, and B[j+1,k+1] is the coefficient of the interaction term X_j Z_k.

norm

The penalty to use for the rows and columns of matrix B. The two possible parameters are "l2" and "l_inf" for the gorup lasso and the infinity norm.

quad

A logical variable indicating if we wish to include quadratic terms when X=Z.

iter

The maximum number of iterations for the ADMM algorithm.

e.abs

An absolute tolerance for convergence.

e.rel

A relative tolerance for convergence. These are used to find a stopping criterion for the ADMM as done in Section 3.3.1 of Boyd, Stephen, et al. 2011

maxiter.B

The maximum number of iterations for updating B via the iterative algorithm for logistic regression.

tol.B

The absolute tolerance for the convergence of B for each iteration of the ADMM algorithm in the case of logistic regression.

verbose

Logical variable which indicates if extra statements should be printed showing progress of the algorithm.

Details

This function fits a regression model based with pair-wise interaction terms by solving the optimization problem (33)(linear regression) or (35)(logistic regression) in Haris, Witten and Simon (2014). The optimization problem is solved via an ADMM algorithm.

Value

The function returns a list where the first component, Estimate, is a list of dimensions length(alphas)*length(lambdas) where $Estimate[[$alpha[a]]][[$lambda[l]]] is an object with components

finB

The estimated coefficient matrix B_est obtained by the ADMM algorithm for minimizing the above objective function.

B, D, E, F

The matrices used in intermediate steps of the ADMM algorithm. We note that numerically all these matrices converge to finB. These matrices are primarily used internally within the function. For details regarding these matrices/notation, we refer the reader to Haris, Witten and Simon (2014).

glist

A list of final estimates for the dual variable of the ADMM algorithm

rho

The last value of the augmented Lagrangian parameter ρ used for the ADMM.

conv

A logical variable stating if the algorithm converged within the maximum number of iterations

iter

The number of iterations for which our algorithm ran. If the algorithm did not converge this will just be equal to the input parameter iter.

The function also returns the training data used to fit the model and the path of penalty parameters for which we estimated the model.

References

Haris, Witten and Simon (2014). Convex Modeling of Interactions with Strong Heredity. Available on ArXiv at http://arxiv.org/abs/1410.3517.

Boyd, Stephen, et al. "Distributed optimization and statistical learning via the alternating direction method of multipliers." Foundations and Trends? in Machine Learning 3.1 (2011): 1-122.

See Also

coef.FAMILY, predict.FAMILY

Examples

  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
library(FAMILY)
library(pROC)
library(pheatmap)

#####################################################################################
#####################################################################################
############################# EXAMPLE - CONTINUOUS RESPONSE #########################
#####################################################################################
#####################################################################################

############################## GENERATE DATA ########################################

#Generate training set of covariates X and Z
set.seed(1)
X.tr<- matrix(rnorm(10*100),ncol = 10, nrow = 100)
Z.tr<- matrix(rnorm(15*100),ncol = 15, nrow = 100)


#Generate test set of covariates X and Z
X.te<- matrix(rnorm(10*100),ncol = 10, nrow = 100)
Z.te<- matrix(rnorm(15*100),ncol = 15, nrow = 100)

#Scale appropiately
meanX<- apply(X.tr,2,mean)
meanY<- apply(Z.tr,2,mean)

X.tr<- scale(X.tr, scale = FALSE)
Z.tr<- scale(Z.tr, scale = FALSE)
X.te<- scale(X.te,center = meanX,scale = FALSE)
Z.te<- scale(Z.te,center = meanY,scale = FALSE)

#Generate full matrix of Covariates
w.tr<- c()
w.te<- c()
X1<- cbind(1,X.tr)
Z1<- cbind(1,Z.tr)
X2<- cbind(1,X.te)
Z2<- cbind(1,Z.te)

for(i in 1:16){
  for(j in 1:11){
    w.tr<- cbind(w.tr,X1[,j]*Z1[,i])
    w.te<- cbind(w.te, X2[,j]*Z2[,i])
  }
}

#Generate response variables with signal from 
#First 5 X features and 5 Z features.

#We construct the coefficient matrix B.
#B[1,1] contains the intercept
#B[-1,1] contains the main effects for X.
# For instance, B[2,1] is the main effect for the first feature in X.
#B[1,-1] contains the main effects for Z.
# For instance, B[1,10] is the coefficient for the 10th feature in Z.
#B[i+1,j+1] is the coefficient of X_i Z_j
B<- matrix(0,ncol = 16,nrow = 11)
rownames(B)<- c("inter" , paste("X",1:(nrow(B)-1),sep = ""))
colnames(B)<- c("inter" , paste("Z",1:(ncol(B)-1),sep = ""))

# First, we simulate data as follows:
# The first five features in X, and the first five features in Z, are non-zero.
# And given the non-zero main effects, all possible interactions are involved.
# We call this "high strong heredity"
B_high_SH<- B
B_high_SH[1:6,1:6]<- 1
#View true coefficient matrix
pheatmap(as.matrix(B_high_SH), scale="none", 
         cluster_rows=FALSE, cluster_cols=FALSE)

Y_high_SH <- as.vector(w.tr%*%as.vector(B_high_SH))+rnorm(100,sd = 2)
Y_high_SH.te <- as.vector(w.te%*%as.vector(B_high_SH))+rnorm(100,sd = 2)

# Now a new setting:
# Again, the first five features in X, and the first five features in Z, are involved. 
# But this time, only a subset of the possible interactions are involved.
# Strong heredity is still maintained. 
# We call this "low strong heredity"
B_low_SH<- B_high_SH
B_low_SH[2:6,2:6]<-0
B_low_SH[3:4,3:5]<- 1
#View true coefficient matrix
pheatmap(as.matrix(B_low_SH), scale="none", 
         cluster_rows=FALSE, cluster_cols=FALSE)
Y_low_SH <- as.vector(w.tr%*%as.vector(B_low_SH))+rnorm(100,sd = 1.5)
Y_low_SH.te <- as.vector(w.te%*%as.vector(B_low_SH))+rnorm(100,sd = 1.5)


############################## FIT SOME MODELS ########################################

#Define alphas and lambdas
#Define 3 different alpha values
#Low alpha values penalize groups more
#High alpha values penalize individual Interactions more
alphas<- c(0.01,0.5,0.99)
lambdas<- seq(0.1,1,length = 50)

#high Strong heredity with l2 norm
fit_high_SH<- FAMILY(X.tr, Z.tr, Y_high_SH, lambdas , 
                     alphas, quad = TRUE,iter=500, verbose = TRUE )
yhat_hSH<- predict(fit_high_SH, X.te, Z.te)
mse_hSH <-apply(yhat_hSH,c(2,3), "-" ,Y_high_SH.te)
mse_hSH<- apply(mse_hSH^2,c(2,3),sum)

#Find optimal model and plot matrix
im<- which(mse_hSH==min(mse_hSH),TRUE)
plot(fit_high_SH$Estimate[[im[2] ]][[im[1]]])


#Plot some matrices for different alpha values
#Low alpha, higher penalty on groups
plot(fit_high_SH$Estimate[[ 1 ]][[ 25 ]])
#Medium alpha, equal penalty on groups and individual interactions
plot(fit_high_SH$Estimate[[ 2 ]][[ 25  ]])
#High alpha, more penalty on individual interactions
plot(fit_high_SH$Estimate[[ 3 ]][[ 40 ]])


#View Coefficients
coef(fit_high_SH)[[im[2]]][[im[1]]]

############################## Uncomment code for EXAMPLE ###########################
# #high Strong heredity with l_infinity norm norm
# fit_high_SH<- FAMILY(X.tr, Z.tr, Y_high_SH, lambdas , 
#                      alphas, quad = TRUE,iter=500, verbose = TRUE,
#                      norm = "l_inf")
# yhat_hSH<- predict(fit_high_SH, X.te, Z.te)
# mse_hSH <-apply(yhat_hSH,c(2,3), "-" ,Y_high_SH.te)
# mse_hSH<- apply(mse_hSH^2,c(2,3),sum)
# 
# #Find optimal model and plot matrix
# im<- which(mse_hSH==min(mse_hSH),TRUE)
# plot(fit_high_SH$Estimate[[im[2] ]][[im[1]]])
# 
# 
# #Plot some matrices for different alpha values
# #Low alpha, higher penalty on groups
# plot(fit_high_SH$Estimate[[ 1 ]][[ 30 ]])
# #Medium alpha, equal penalty on groups and individual interactions
# plot(fit_high_SH$Estimate[[ 2 ]][[ 10 ]])
# #High alpha, more penalty on individual interactions
# plot(fit_high_SH$Estimate[[ 3 ]][[ 20 ]])
# 
# 
# #View Coefficients
# coef(fit_high_SH)[[im[2]]][[im[1]]]


############################## Uncomment code for EXAMPLE ###########################
# #Redefine lambdas
# lambdas<- seq(0.1,0.5,length = 50)
# 
# #low Strong heredity with l_2 norm
# fit_low_SH<- FAMILY(X.tr, Z.tr, Y_low_SH, lambdas , 
#                      alphas, quad = TRUE,iter=500, verbose = TRUE )
# yhat_lSH<- predict(fit_low_SH, X.te, Z.te)
# mse_lSH <-apply(yhat_lSH,c(2,3), "-" ,Y_low_SH.te)
# mse_lSH<- apply(mse_lSH^2,c(2,3),sum)
# 
# #Find optimal model and plot matrix
# im<- which(mse_lSH==min(mse_lSH),TRUE)
# plot(fit_low_SH$Estimate[[im[2] ]][[im[1]]])
# 
# 
# #Plot some matrices for different alpha values
# #Low alpha, higher penalty on groups
# plot(fit_low_SH$Estimate[[ 1 ]][[ 25 ]])
# #Medium alpha, equal penalty on groups and individual interactions
# plot(fit_low_SH$Estimate[[ 2 ]][[ 10 ]])
# #High alpha, more penalty on individual interactions
# plot(fit_low_SH$Estimate[[ 3 ]][[ 10 ]])
# 
# 
# #View Coefficients
# coef(fit_low_SH)[[im[2]]][[im[1]]]


#####################################################################################
#####################################################################################
############################### EXAMPLE - BINARY RESPONSE ###########################
#####################################################################################
#####################################################################################

############################## GENERATE DATA ########################################

#Generate data for logistic regression
Yp_high_SH<- as.vector((w.tr)%*%as.vector(B_high_SH))
Yp_high_SH.te<- as.vector((w.te)%*%as.vector(B_high_SH))

Yprobs_high_SH<- 1/(1+exp(-Yp_high_SH))
Yprobs_high_SH.te<- 1/(1+exp(-Yp_high_SH.te))

Yp_high_SH<- rbinom(100, size = 1, prob = Yprobs_high_SH)
Yp_high_SH.te<- rbinom(100, size = 1, prob = Yprobs_high_SH.te)

lambdas<- seq(0.01,0.15,length = 50)

############################## FIT SOME MODELS ########################################

#Fit glm via l_2 norm
fit_high_SH<- FAMILY(X.tr, Z.tr, Yp_high_SH, lambdas , 
                    alphas, quad = TRUE,iter=500, verbose = TRUE,
                    family = "binomial")
yhp_hSH<- predict(fit_high_SH, X.te, Z.te)
mse_high_SH <-apply(yhp_hSH,c(2,3), "-" ,Yp_high_SH.te)
mse_hSH<- apply(mse_high_SH^2,c(2,3),sum)
im<- which(mse_hSH==min(mse_hSH),TRUE)
plot(fit_high_SH$Estimate[[im[2] ]][[im[1]]])
roc( Yp_high_SH.te,yhp_hSH[,im[1],im[2]],plot = TRUE)

#View Coefficients
coef(fit_high_SH)[[im[2]]][[im[1]]]

############################## Uncomment code for EXAMPLE ###########################
# #Fit glm via l_infinity norm
# fit_high_SH<- FAMILY(X.tr, Z.tr, Yp_high_SH, lambdas , norm = "l_inf",
#                      alphas, quad = TRUE,iter=500, verbose = TRUE,
#                      family = "binomial")
# yhp_hSH<- predict(fit_high_SH, X.te, Z.te)
# mse_high_SH <-apply(yhp_hSH,c(2,3), "-" ,Yp_high_SH.te)
# mse_hSH<- apply(mse_high_SH^2,c(2,3),sum)
# im<- which(mse_hSH==min(mse_hSH),TRUE)
# plot(fit_high_SH$Estimate[[im[2] ]][[im[1]]])
# roc( Yp_high_SH.te,yhp_hSH[,im[1],im[2]],plot = TRUE)
# 
# #View Coefficients
# coef(fit_high_SH)[[im[2]]][[im[1]]]

#####################################################################################
#####################################################################################
############################## EXAMPLE WHERE X=Z #################################### 
######################## Uncomment Code for EXAMPLE #################################
#####################################################################################

############################## GENERATE DATA ########################################
# #Redefine Lambdas
# lambdas<- seq(0.01,0.3,length = 50)
# 
# 
# #We consider the case X=Z now
# w.tr<- c()
# w.te<- c()
# X1<- cbind(1,X.tr)
# X2<- cbind(1,X.te)
# 
# for(i in 1:11){
#   for(j in 1:11){
#     w.tr<- cbind(w.tr,X1[,j]*X1[,i])
#     w.te<- cbind(w.te, X2[,j]*X2[,i])
#   }
# }
# 
# B<- matrix(0,ncol = 11,nrow = 11)
# rownames(B)<- c("inter" , paste("X",1:(nrow(B)-1),sep = ""))
# colnames(B)<- c("inter" , paste("X",1:(ncol(B)-1),sep = ""))
# 
# 
# B_high_SH<- B
# B_high_SH[1:6,1:6]<- 1
# #We exclude quadratic terms in this example
# diag(B_high_SH)[-1]<-0
# #View true coefficient matrix
# pheatmap(as.matrix(B_high_SH), scale="none", 
#          cluster_rows=FALSE, cluster_cols=FALSE)
# 
# #With high Strong heredity: all possible interactions
# Y_high_SH <- as.vector(w.tr%*%as.vector(B_high_SH))+rnorm(100)
# Y_high_SH.te <- as.vector(w.te%*%as.vector(B_high_SH))+rnorm(100)
# 
# ############################## FIT SOME MODELS ########################################
# 
# #high Strong heredity with l_2 norm
# fit_high_SH<- FAMILY(X.tr, X.tr, Y_high_SH, lambdas , 
#                      alphas, quad = FALSE,iter=500, verbose = TRUE )
# yhat_hSH<- predict(fit_high_SH, X.te, X.te)
# mse_hSH <-apply(yhat_hSH,c(2,3), "-" ,Y_high_SH.te)
# mse_hSH<- apply(mse_hSH^2,c(2,3),sum)
# 
# #Find optimal model and plot matrix
# im<- which(mse_hSH==min(mse_hSH),TRUE)
# plot(fit_high_SH$Estimate[[im[2] ]][[im[1]]])
# 
# 
# #Plot some matrices for different alpha values
# #Low alpha, higher penalty on groups
# plot(fit_high_SH$Estimate[[ 1 ]][[ 50 ]])
# #Medium alpha, equal penalty on groups and individual interactions
# plot(fit_high_SH$Estimate[[ 2 ]][[ 50 ]])
# #High alpha, more penalty on individual interactions
# plot(fit_high_SH$Estimate[[ 3 ]][[ 50 ]])
# 
# 
# #View Coefficients
# coef(fit_high_SH,XequalZ = TRUE)[[im[2]]][[im[1]]]

FAMILY documentation built on May 30, 2017, 3:10 a.m.