多类分类图？

Question

I was looking through the scikit-learn documentation and realized that most of the plotting curve support exists fro binary classification.

I was wanting to plot the precision_recall curve and the learning curve for my multi-class classifier.

model1=LogisticRegression()
y_d = model1.predict_proba(matrix_test)

I was wondering if there exists any method for plotting the precision_recall curve and learning curve for my classifer with a sparse matrix_test of size (22428,22000) and labels being a np array of size (22428,)?

Answer 1

If you look at the definiton of Precision and Recall , you can see that there is an asymmetry there that does not directly translation into higher dimensions. Say the classes are C ₁ and C ₂ , then arbitrarily , one of them is considered "True", and the other "False" (you can also notice that there is no symmetry: reversing "True" and "False" will not give the same results). In higher dimensions, this simply can't be done directly.

There are many ways to come up with heuristic extensions, though. Say your classes are C ₁ , ..., C _m . You can calculate m precisions and recalls from the view point of class C _i , then take a (weighted) average. The weights should probably reflect the importance of the classes.

Note that this is exactly the scheme used for the binary case, where the weight for the "True" class is chosen to be 1, and the weight for the "False" class is chosen to be 0 (again, emphasizing the arbitrary asymmetry of this score).

In terms of implementation, this is trivial. Say your confusion matrix is m . Then from the point of view of class i , the precision is m[i, i] / np.sum(m[:, i]) and the recall is m[i, i] / np.sum(m[i, :]) .