C4.5 ID3算法计算的是信息增益，C4.5算法计算的是信息增益比，对上面ID3版本的函数稍作修改即可
(self,X,y)
| 97 | return bestFeatureIndex |
| 98 | |
| 99 | def _chooseBestFeatureToSplit_C45(self,X,y): |
| 100 | """C4.5 |
| 101 | ID3算法计算的是信息增益，C4.5算法计算的是信息增益比，对上面ID3版本的函数稍作修改即可 |
| 102 | """ |
| 103 | numFeatures = X.shape[1] |
| 104 | oldEntropy = self._calcEntropy(y) |
| 105 | bestGainRatio = 0.0 |
| 106 | bestFeatureIndex = -1 |
| 107 | #对每个特征都计算一下gainRatio=infoGain/splitInformation |
| 108 | for i in range(numFeatures): |
| 109 | featList = X[:,i] |
| 110 | uniqueVals = set(featList) |
| 111 | newEntropy = 0.0 |
| 112 | splitInformation = 0.0 |
| 113 | #对第i个特征的各个value，得到各个子数据集，计算各个子数据集的熵， |
| 114 | #进一步地可以计算得到根据第i个特征分割原始数据集后的熵newEntropy |
| 115 | for value in uniqueVals: |
| 116 | sub_X,sub_y = self._splitDataSet(X,y,i,value) |
| 117 | prob = len(sub_y)/float(len(y)) |
| 118 | newEntropy += prob * self._calcEntropy(sub_y) |
| 119 | splitInformation -= prob * np.log2(prob) |
| 120 | #计算信息增益比，根据信息增益比选择最佳分割特征 |
| 121 | #splitInformation若为0，说明该特征的所有值都是相同的，显然不能作为分割特征 |
| 122 | if splitInformation==0.0: |
| 123 | pass |
| 124 | else: |
| 125 | infoGain = oldEntropy - newEntropy |
| 126 | gainRatio = infoGain/splitInformation |
| 127 | if(gainRatio > bestGainRatio): |
| 128 | bestGainRatio = gainRatio |
| 129 | bestFeatureIndex = i |
| 130 | return bestFeatureIndex |
| 131 | |
| 132 | |
| 133 |
no test coverage detected