bmi = values.index(max(values)) # Best Model Index
q2, NV, NVL_Model, NVL_OPS = float(results[bmi][0]), int(results[bmi][1]), int(
results[bmi][2]), int(results[bmi][3])
if len(sys.argv) > 5:
start = 4
else:
infoVec = int(results[bmi][4])
start = 5
sel = []
for i in range(start, len(results[bmi])):
sel.append(int(results[bmi][i]))
sel = np.array(sel)
q2, r2, rmsecv, rmse, corrcv, corrmdl = cv.plsLOO(X[:, sel], Y, int(NVL_Model))
if len(sys.argv) > 5:
output = np.array([
q2, r2, rmsecv, rmse, corrcv, corrmdl,
int(NV),
int(NVL_Model),
int(NVL_OPS)
])
print('Best model found:', 'Q2:', '{0:.4f}'.format(q2), 'NV:', int(NV),
'NVL_Model:', int(NVL_Model), 'NVL_OPS:', int(NVL_OPS))
else:
output = np.array([
q2, r2, rmsecv, rmse, corrcv, corrmdl,
int(NV),
int(NVL_Model),
def OPS(X, Y, maxVL_OPS, maxVL_Model, maxVariables, infoVec='prod', verbose=0):
"""
Ordered Predictors Selection (OPS): feature selection algorithm for multivariate regression.
For details about OPS, please see reference [1].
Parameters
----------
X : numpy array of shape [n_samples,n_features]
Samples.
Y : numpy array of shape [n_samples]
Target values.
maxVL_OPS: integer
Maximum number of latent variables tested in OPS for sorting varables (when using 'prod' or 'reg' infoVec).
maxVL_Model: integer
Maximum number of latent variables tested in PLS models with each feature subset.
maxVariables: integer
Maximum number of features tested.
infoVec : {'prod', 'reg', or 'corr'}
The informative vector used to sort variables.
'corr' is the correlation vector between features and target values,
'reg' is the regression vector obtained with a PLS model, and
'prod' is the combination of correlation vector and regression vector (usually is best)
verbose : {0,1}
0 Verbosity mode off.
1 Verbosity mode on.
Returns
-------
subset : integer list
Index columns of selected features.
q2 : float
Q^2 external validation metric. For details see reference [2].
NV : int
Number of selected features.
NVL_Model: float
Number of PLS latent variables.
NVL_OPS : int
Number of OPS latent variables used for sorting variables.
References
----------
[1] Teófilo, R. F.; Martins, J. P. A. & Ferreira, M. M. C. (2009).
Sorting variables by using informative vectors as a strategy for feature selection in multivariate regression.
Journal of Chemometrics, 23(1):32--48. ISSN 0886-9383
[2] Martins, J. P. A.; Barbosa, E. G.; Pasqualoto, K. F. M. & Ferreira, M. M. C. (2009).
LQTA-QSAR: a new 4D-QSAR methodology.
Journal of Chemical Information and Modeling, 49(6):1428--1436. ISSN 1549-9596
"""
nSamples, nFeatures = X.shape
if maxVariables > nFeatures:
maxVariables = nFeatures
# Best models are saved in this matrix, each line is formatted as
# [Q^2, NumVariables, numLatentVariables (Model), NumLatentVariables (OPS), Key of features subset]
results = np.empty([0, 5])
selFeats = {}
key_features = 1
if infoVec == 'corr':
maxVL_OPS = 1
for NVL_OPS in range(1, maxVL_OPS + 1):
if verbose == 1:
stdout = 'Running OPS ' + str(int(
(NVL_OPS / maxVL_OPS) * 100)) + '%/' + str(
int((maxVL_OPS / maxVL_OPS) * 100)) + '%'
print(stdout)
iterResults = np.empty([0, 5])
iterSelFeats = {}
if infoVec == 'corr':
corrVec = correlationVector(X, Y)
informativeVector = formatInfoVec(corrVec, nFeatures)
elif infoVec == 'reg':
regVec = regressionVector(X, Y, NVL_OPS)
informativeVector = formatInfoVec(regVec, nFeatures)
else:
corrVec = correlationVector(X, Y)
regVec = regressionVector(X, Y, NVL_OPS)
prodVec = regVec * corrVec
informativeVector = formatInfoVec(prodVec, nFeatures)
_, mapping = sortVariables(X, Y, informativeVector)
for NV in range(1, maxVariables + 1):
subset = mapping[0:NV]
for NVL_Model in range(1, maxVL_Model + 1):
if NV < NVL_Model:
continue
q2, _, _, _, _, _ = cv.plsLOO(X[:, subset], Y, NVL_Model)
metrics = np.array([q2, NV, NVL_Model, NVL_OPS, key_features])
iterSelFeats[key_features] = subset
key_features = key_features + 1
iterResults = np.vstack((iterResults, metrics))
bestModelIndex = np.argmax(iterResults[:, 0])
bestModel = iterResults[bestModelIndex, :]
if verbose == 1:
print('Iteration\'s best model:',
'Q2:', '{0:.4f}'.format(bestModel[0]), 'NV:',
int(bestModel[1]), 'NVL_Model:', int(bestModel[2]),
'NVL_OPS:', int(bestModel[3]))
results = np.vstack((results, bestModel))
selFeats[bestModel[4]] = iterSelFeats[bestModel[4]]
maxPerf = np.argmax(results[:, 0])
q2 = results[maxPerf, 0]
NV = results[maxPerf, 1]
NVL_Model = results[maxPerf, 2]
NVL_OPS = results[maxPerf, 3]
subset = selFeats[results[maxPerf, 4]]
if verbose == 1:
print('Best model found:', 'Q2:', '{0:.4f}'.format(q2), 'NV:', int(NV),
'NVL_Model:', int(NVL_Model), 'NVL_OPS:', int(NVL_OPS))
return subset, q2, NV, NVL_Model, NVL_OPS
def OPS_auto(X, Y, NVL_OPS, maxVL_Model, maxVariables, verbose=0):
"""
Automated version of Ordered Predictors Selection (OPS):
feature selection algorithm for multivariate regression. For details about OPS, please see reference [1].
This implemetation automats the process testing all available informative vectors for sorting variables.
It's runs slowly than the previous implementation.
Parameters
----------
X : numpy array of shape [n_samples,n_features]
Samples.
Y : numpy array of shape [n_samples]
Target values.
NVL_OPS: integer
Maximum number of latent variables tested in OPS for sorting varables (when using 'prod' or 'reg' infoVec).
maxVL_Model: integer
Maximum number of latent variables tested in PLS models with each feature subset.
maxVariables: integer
Maximum number of features tested.
verbose : {0,1}
0 Verbosity mode off.
1 Verbosity mode on.
Returns
-------
subset : integer list
Index columns of selected features.
q2 : float
Q^2 external validation metric. For details see reference [2].
NV : int
Number of selected features.
NVL_Model: float
Number of PLS latent variables.
NVL_OPS : int
Number of OPS latent variables used for sorting variables.
infoVec : int
Returns the informative vector used for sorting variables in the best model found.
1 if it's correlation vector.
2 if it's regression vector.
3 if it's combination of correlation and regression vectors
References
----------
[1] Teófilo, R. F.; Martins, J. P. A. & Ferreira, M. M. C. (2009).
Sorting variables by using informative vectors as a strategy for feature selection in multivariate regression.
Journal of Chemometrics, 23(1):32--48. ISSN 0886-9383
[2] Martins, J. P. A.; Barbosa, E. G.; Pasqualoto, K. F. M. & Ferreira, M. M. C. (2009).
LQTA-QSAR: a new 4D-QSAR methodology.
Journal of Chemical Information and Modeling, 49(6):1428--1436. ISSN 1549-9596
"""
nSamples, nFeatures = X.shape
if maxVariables > nFeatures:
maxVariables = nFeatures
# Best models are saved in this matrix, each line is formatted as
# [Q^2, NumVariables, numLatentVariables (Model), NumLatentVariables (OPS), Key of features subset, Informative Vector]
results = np.empty([0, 6])
selFeats = {}
key_features = 1
iterResults = np.empty([0, 6])
iterSelFeats = {}
regVec = regressionVector(X, Y, NVL_OPS)
corrVec = correlationVector(X, Y)
prodVec = regVec * corrVec
# Begin testing with correlation vetor
informativeVector = formatInfoVec(corrVec, nFeatures)
_, mapping = sortVariables(X, Y, informativeVector)
for NV in range(1, maxVariables + 1):
subset = mapping[0:NV]
for NVL_Model in range(1, maxVL_Model + 1):
if NV < NVL_Model:
continue
q2, _, _, _, _, _ = cv.plsLOO(X[:, subset], Y, NVL_Model)
metrics = np.array([q2, NV, NVL_Model, NVL_OPS, key_features, 1])
iterSelFeats[key_features] = subset
key_features = key_features + 1
iterResults = np.vstack((iterResults, metrics))
# End testing with correlation vetor
# Begin testing with regression vetor
informativeVector = formatInfoVec(regVec, nFeatures)
_, mapping = sortVariables(X, Y, informativeVector)
for NV in range(1, maxVariables + 1):
subset = mapping[0:NV]
for NVL_Model in range(1, maxVL_Model + 1):
if NV < NVL_Model:
continue
q2, _, _, _, _, _ = cv.plsLOO(X[:, subset], Y, NVL_Model)
metrics = np.array([q2, NV, NVL_Model, NVL_OPS, key_features, 2])
iterSelFeats[key_features] = subset
key_features = key_features + 1
iterResults = np.vstack((iterResults, metrics))
# End testing with regression vetor
# Begin testing with product vetor
informativeVector = formatInfoVec(prodVec, nFeatures)
_, mapping = sortVariables(X, Y, informativeVector)
for NV in range(1, maxVariables + 1):
subset = mapping[0:NV]
for NVL_Model in range(1, maxVL_Model + 1):
if NV < NVL_Model:
continue
q2, _, _, _, _, _ = cv.plsLOO(X[:, subset], Y, NVL_Model)
metrics = np.array([q2, NV, NVL_Model, NVL_OPS, key_features, 3])
iterSelFeats[key_features] = subset
key_features = key_features + 1
iterResults = np.vstack((iterResults, metrics))
# End testing with product vetor
bestModelIndex = np.argmax(iterResults[:, 0])
bestModel = iterResults[bestModelIndex, :]
if verbose == 1:
print('Iteration\'s best model:', 'Q2:', '{0:.4f}'.format(bestModel[0]), 'NV:', int(bestModel[1]), 'NVL_Model:', int(bestModel[2]), 'NVL_OPS:', int(bestModel[3]), 'InfoVec', int(bestModel[5]))
results = np.vstack((results, bestModel))
selFeats[bestModel[4]] = iterSelFeats[bestModel[4]]
maxPerf = np.argmax(results[:, 0])
q2 = results[maxPerf, 0]
NV = results[maxPerf, 1]
NVL_Model = results[maxPerf, 2]
NVL_OPS = results[maxPerf, 3]
infoVec = results[maxPerf, 5]
subset = selFeats[results[maxPerf, 4]]
return subset, q2, NV, NVL_Model, NVL_OPS, infoVec