Classification

Consider a binary classification problem. The data and target files are available online. The domain of the problem is chemoinformatics. Data is about toxicity of 4K small molecules. The creation of a predictive system happens in 3 steps:

data conversion: transform instances into a suitable graph format. This is done using specialized programs for each (domain, format) pair. In the example we have molecular graphs encoded using the gSpan format and we will therefore use the 'gspan' tool.
data vectorization: transform graphs into sparse vectors. This is done using the EDeN tool. The vectorizer accepts as parameters the (maximal) size of the fragments to be used as features, this is expressed as the pair 'radius' and the 'distance'. See for details: F. Costa, K. De Grave,''Fast Neighborhood Subgraph Pairwise Distance Kernel'', 27th International Conference on Machine Learning (ICML), 2010.
modelling: fit a predicitve system and evaluate its performance. This is done using the tools offered by the scikit library. In the example we will use a Stochastic Gradient Descent linear classifier.

In the following cells there is the code for each step.

Install the library

pip install git+https://github.com/fabriziocosta/EDeN.git --user

1 Conversion

load a target file



In [1]:

    
from eden.util import load_target
y = load_target( 'http://www.bioinf.uni-freiburg.de/~costa/bursi.target' )

load data and convert it to graphs



In [2]:

    
from eden.converter.graph.gspan import gspan_to_eden
graphs = gspan_to_eden( 'http://www.bioinf.uni-freiburg.de/~costa/bursi.gspan' )

2 Vectorization

setup the vectorizer



In [3]:

    
from eden.graph import Vectorizer
vectorizer = Vectorizer( r=2,d=5 )

extract features and build data matrix



In [4]:

    
%%time
X = vectorizer.transform( graphs )
print 'Instances: %d Features: %d with an avg of %d features per instance' % (X.shape[0], X.shape[1],  X.getnnz()/X.shape[0])









    



Instances: 4337 Features: 1048577 with an avg of 167 features per instance
CPU times: user 44.8 s, sys: 373 ms, total: 45.2 s
Wall time: 46.9 s

3 Modelling

Induce a predictor and evaluate its performance



In [5]:

    
%%time
#induce a predictive model
from sklearn.linear_model import SGDClassifier
predictor = SGDClassifier(average=True, class_weight='auto', shuffle=True, n_jobs=-1)

from sklearn import cross_validation
scores = cross_validation.cross_val_score(predictor, X, y, cv=10, scoring='roc_auc')

import numpy as np
print('AUC ROC: %.4f +- %.4f' % (np.mean(scores),np.std(scores)))









    



AUC ROC: 0.9071 +- 0.0111
CPU times: user 1.09 s, sys: 176 ms, total: 1.27 s
Wall time: 1.28 s