Local sparsity control for Naive Bayes with extreme misclassiication costs

[flrngel]

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.66.5667&rep=rep1&type=pdf

1. Introduction

• In text domain, there is excessive number of features
• To control sparsity, "threshold to cut-off feature" was used traditionally
• This paper suggests that local approaches (feature selection) has potential benefit

4. Sparsity control via feature selection

• Global sparsity cut-off (feature ranking) is better than feature count cut-off
• Cannot say local approach is always better than global approach but seems to better on many cases

6. Datasets

6.2. Model comparision

• NBLOC is best

7. Results

8. Conclusions

• standard Naive Bayes classifier has propensity to make errors with high confidence
  • especially in the text domain where overconfidence can come from large dimensionality of the feature
• paper claims to use local approach and document-specific approach
• local feature selection may preferable which dataset and feature ranking functions are considered
• Naive Bayes could perform better with document-specific feature selection at cost settings
• paper shows Naive Bayes with document length normalization and TFIDF term weighting makes benefit