===================================== Multi (& Mono) Data-Driven Sparse PLS

*mddspls is the python light package of the data-driven sparse PLS algorithm*

In the high dimensional settings (large number of variables), one objective is to select the relevant variables and thus to reduce the dimension. That subspace selection is often managed with supervised tools. However, some data can be missing, compromising the validity of the sub-space selection. We propose a PLS, Partial Least Square, based method, called dd-sPLS for data-driven-sparse PLS, allowing jointly variable selection and subspace estimation while training and testing missing data imputation through a new algorithm called Koh-Lanta.

It contains one main class mddspls and one associated important method denote predict permitting to predict from a new dataset. The function called perf_mddsPLS permits to compute cross-validation.

Data simulation

One might be interested to simulate data and test the package through regression and classification. Here a spiked model is used::

#!/usr/bin/env python

import py_ddspls
import numpy as np
import sklearn.metrics as sklm

n = 100
R_model = 10
R_X_Y = 2
T = 10
L = np.array(np.random.normal(0,1,n*R_model)).reshape((n,R_model))

p_t = 20
q = 5

Xs = {}

for t in range(T):
    Omega_1_2 = np.diag(np.random.uniform(0,1,R_model))
    u,s,vh = np.linalg.svd(np.array(np.random.normal(0,1,p_t*p_t)).reshape((p_t,p_t)))
    U_mod_T = vh[0:R_model,:]
    Xs[t] = L@Omega_1_2@U_mod_T

Omega_y_1_2 = np.diag(np.concatenate((np.ones(R_X_Y),np.zeros(R_model-R_X_Y))))
u,s,vh = np.linalg.svd(np.array(np.random.normal(0,1,R_model*R_model)).reshape((R_model,R_model)))
U_mod_T = vh[:,0:q]
Y = L@Omega_y_1_2@U_mod_T

k_groups = 2
Y_transfor = Y[:,0]
lolo = np.linspace(np.min(Y_transfor),np.max(Y_transfor),k_groups+1)
Y_bin = np.zeros(n)
for ii in range(n):
    for k_i in range(k_groups):
        if (Y_transfor[ii]>=lolo[k_i])&(Y_transfor[ii]<lolo[k_i+1]):
            Y_bin[ii] = k_i
        if Y_transfor[ii]==lolo[k_groups]:
            Y_bin[ii] = k_groups-1

pos_0 = np.where(Y_bin==0)[0]
pos_1 = np.where(Y_bin==1)[0]
Y_classif = np.repeat("Class 2",n)
Y_classif[pos_1] = "Class 1"

# Missing values are introduced in blocks 1, 2 and 3
Xs[0][0,:] = None
Xs[1][1:3,:] = None
Xs[2][2:10,:] = None

The dd-sPLS regularization parameter is fixed to 0.6::

lambd=0.6

A train/test dataset is defined for the sack of the example::

id_train = range(30,100)
id_test = range(30)
Xtrain = {}
Ytrain = Y[id_train,:]
Xtest = {}

Regression analysis

Let us produce 2 axes::

R=2

Start model building and check results with sklearn tools::

mod_0=py_ddspls.model.ddspls(Xtrain,Ytrain,lambd=lambd,R=R,mode="reg",verbose=True)
Y_est_reg = mod_0.predict(Xtest)        
print(sklm.mean_squared_error(Y[id_test,:],Y_est_reg))

Leave-one-out cross validation can be performed with built tools, the parameter NCORES permits to fix the number of cores to be used in the process ::

perf_model_reg = py_ddspls.model.perf_ddspls(Xs,Y,R=R,kfolds="loo",n_lambd=10,NCORES=4,mode="reg")
print(perf_model_reg)

fig = plt.figure()
ax = fig.add_subplot(1, 1, 1)
cols = ['r','g','b','y','black','brown']
for jj in range(q):
    ax.plot(perf_model_reg[:,1],perf_model_reg[:,jj+2],cols[jj])

ax.plot(perf_model_reg[:,1],
    np.sqrt((perf_model_reg[:,2:(2+q)]**2).mean(axis=1)),"brown",linewidth=2,ls="--")
plt.legend(np.concatenate((1+np.arange(q),np.array("RMSE of RMSE errors").reshape((1,)))),loc='upper')
plt.title('Leave-One-Out Cross-validation error against $\lambda$')
plt.xlabel('$\lambda$')
plt.ylabel('RMSE')
plt.show()

Which returns this kind of graphics

.. image:: https://raw.githubusercontent.com/hlorenzo/py_ddspls/master/images/reg.png :width: 600

For 0.9 one can find a minimum of the RMSE of the RMSE of each variable. This oservation can be mitigated assuming that only Y variables 1 and 4 are well described by the X dataset. In that context, a discussion with experts, might help to decide the value to give to the parameter.

Classification analysis

Let us produce 1 axis since only one group must be discriminated::

R=1

Start model building and tcheck results with sklearn tools::

mod_0_classif=py_ddspls.model.ddspls(Xs,Y_bin,lambd=lambd,R=R,mode="clas",verbose=True)
Y_est = mod_0_classif.predict(Xtest)
print(sklm.classification_report(Y_est, Y_classif[id_test]=='Class 1'))

Cross validation can be performed with built tools, the parameter NCORES permits to use parallellization::

perf_model_class = py_ddspls.model.perf_ddspls(Xs,Y_classif,R=R,kfolds="loo",n_lambd=40,NCORES=7,mode="classif")
print(perf_model_class)

fig = plt.figure()
ax = fig.add_subplot(1, 1, 1)
ax.plot(perf_model_class[:,1], perf_model_class[:,2])
plt.title('Leave-One-Out Cross-validation error against $\lambda$')
plt.xlabel('$\lambda$')
plt.ylabel('Classification Error')
plt.show()

Which returns this kind of graphics

.. image:: https://raw.githubusercontent.com/hlorenzo/py_ddspls/master/images/cla.png :width: 600

One that figure one can see that a parameter approximately equal to 0.45 can be chosen.

Enjoy :)