Accurancy of Naive Bayesian classifier?

Question

Is there any percent for accurancy of Naive Bayesian classifier that can we use to check how accurate is our classifier?

See [Cross validation](http://en.wikipedia.org/wiki/Cross-validation_%28statistics%29#K-fold_cross-validation) — Flexo, Aug 02 '12 at 10:28

Johan · Answer 1 · 2012-08-02T10:33:42.157

2

Normally in machine learning one would look at specificity vs. sensitivity to access the performance of the classifier. http://en.wikipedia.org/wiki/Sensitivity_and_specificity

Since there is normally a trade of between true positives, false positives, true negative and false positives, it is important to decide what is more important in the particular application that you are looking at.

This coupled with cross-validation and possibly prediction on a dataset where you have randomized of the decision variable should give you a picture of how good your classifier is.

Remember that there is no such thing as a simple answer of predictor performance. For example, you want to classify a data set where 99 of the samples are of class A, and 1 of them are of class B. A classifier build to classify all of the examples as A would have a 99% sensitivity, but still not be very useful.

edited Aug 02 '12 at 10:33

answered Aug 02 '12 at 10:26

Johan

689
7
17

Does the number of samples in training set in the classes should be almost the same?Does this affect on the accurancy of the classifier? – vikifor Aug 02 '12 at 10:39
It will probably increase the power of the classifier, yes. If they are of extremely different sizes it might be wise to either down sample the large one, or bootstrap the smaller one. For an in-depth discussion of these kinds of concepts i would recommend this book: http://www-stat.stanford.edu/~tibs/ElemStatLearn/ (which is freely available on the linked page) – Johan Aug 02 '12 at 10:45

score 2 · Answer 2 · answered Aug 09 '12 at 00:14

Here is another advice you can use. In information retrieval , F-score is a common used evaluation criterion, where F-score contains two factor. 1. recall = tp / (tp + fn) 2. precision = tp / (tp+fp)

(tp = true positive...etc)

F-score = (2*recall*presicion) / (recall + precision)

Just as Johan said, evaluation criterion vary with different situation. For example , in multi-label problems , some may also consider Hamming Loss or Ranking Loss. I think in most of the single-label case, F-score is the most popular one.

Accurancy of Naive Bayesian classifier?

2 Answers2