Source Code for Module ML.Composite.AdjustComposite

 1  # $Id: AdjustComposite.py 778 2008-07-30 09:51:55Z glandrum $ 
 2  # 
 3  #  Copyright (C) 2003  greg Landrum and Rational Discovery LLC 
 4  #   All Rights Reserved 
 5  # 
 6  """ functionality to allow adjusting composite model contents 
 7   
 8  """ 
 9  import numpy 
10  import copy 
11   
12 -def BalanceComposite(model,set1,set2,weight,targetSize,names1=None,names2=None): 
13    """ adjusts the contents of the composite model so as to maximize 
14      the weighted classification accuracty across the two data sets. 
15   
16      The resulting composite model, with _targetSize_ models, is returned. 
17   
18      **Notes**: 
19   
20        - if _names1_ and _names2_ are not provided, _set1_ and _set2_ should 
21          have the same ordering of columns and _model_ should have already 
22          have had _SetInputOrder()_ called. 
23   
24    """ 
25    # 
26    # adjust the weights to be proportional to the size of the two data sets 
27    #   The normalization we do here assures that a perfect model contributes 
28    #   a score of S1+S2 to the final 
29    # 
30    S1 = len(set1) 
31    S2 = len(set2) 
32    weight1 = float(S1+S2)*(1-weight)/S1 
33    weight2 = float(S1+S2)*weight/S2 
34    #print '\t:::',S1,S2,weight1,weight2 
35    #print 'nModels:',len(model) 
36    # start with a copy so that we get all the additional schnick-schnack 
37    res = copy.copy(model) 
38    res.modelList = [] 
39    res.errList = [] 
40    res.countList = [] 
41    res.quantizationRequirements = [] 
42   
43    startSize = len(model) 
44    scores = numpy.zeros(startSize,numpy.float) 
45    actQuantBounds = model.GetActivityQuantBounds() 
46    if names1 is not None: 
47      model.SetInputOrder(names1) 
48    for pt in set1: 
49      pred,conf = model.ClassifyExample(pt) 
50      if actQuantBounds: 
51        ans = model.QuantizeActivity(pt)[-1] 
52      else: 
53        ans = pt[-1] 
54      votes = model.GetVoteDetails() 
55      for i in range(startSize): 
56        if votes[i]==ans: scores[i] += weight1 
57    if names2 is not None: 
58      model.SetInputOrder(names2) 
59    for pt in set2: 
60      pred,conf = model.ClassifyExample(pt) 
61      if actQuantBounds: 
62        ans = model.QuantizeActivity(pt)[-1] 
63      else: 
64        ans = pt[-1] 
65      votes = model.GetVoteDetails() 
66      for i in range(startSize): 
67        if votes[i]==ans: scores[i] += weight2 
68    # normalize the scores 
69    nPts = S1+S2 
70    scores /= nPts 
71    # sort them: 
72    bestOrder = list(numpy.argsort(scores)) 
73    bestOrder.reverse() 
74    print '\tTAKE:',bestOrder[:targetSize] 
75    # and now take the best set: 
76    for i in range(targetSize): 
77      idx = bestOrder[i] 
78      mdl = model.modelList[idx] 
79      res.modelList.append(mdl) 
80      res.errList.append(1.-scores[idx]) 
81      res.countList.append(1) 
82      # FIX: this should probably be more general: 
83      res.quantizationRequirements.append(0) 
84    return res 
85