A maximum entropy approach to natural language processing
Computational Linguistics
Exploiting generative models in discriminative classifiers
Proceedings of the 1998 conference on Advances in neural information processing systems II
SVM binary classifier ensembles for image classification
Proceedings of the tenth international conference on Information and knowledge management
Transductive Inference for Text Classification using Support Vector Machines
ICML '99 Proceedings of the Sixteenth International Conference on Machine Learning
Kernel Based Image Classification
ICANN '01 Proceedings of the International Conference on Artificial Neural Networks
Image Indexing Using Color Correlograms
CVPR '97 Proceedings of the 1997 Conference on Computer Vision and Pattern Recognition (CVPR '97)
An Introduction to the Conjugate Gradient Method Without the Agonizing Pain
An Introduction to the Conjugate Gradient Method Without the Agonizing Pain
Support vector machines for histogram-based image classification
IEEE Transactions on Neural Networks
Learning from labeled features using generalized expectation criteria
Proceedings of the 31st annual international ACM SIGIR conference on Research and development in information retrieval
Generalized Expectation Criteria for Semi-Supervised Learning with Weakly Labeled Data
The Journal of Machine Learning Research
| Hi-index | 0.00 |
Discriminative and generative methods provide two distinct approaches to machine learning classification. One advantage of generative approaches is that they naturally model the prior class distributions. In contrast, discriminative approaches directly model the conditional distribution of class given inputs, so the class priors are only implicitly obtained if the input density is known. In this paper, we propose a framework for incorporating class prior proportions into discriminative methods in order to improve their classification accuracy. The basic idea is to enforce that the distribution of class labels predicted on the test data by the discriminative model is consistent with the class priors. Therefore, the discriminative model has to not only fit the training data well but also predict class labels for the test data that are consistent with the class priors. Experiments on five different UCI datasets and one image database show that this framework is effective in improving the classification accuracy when the training data and the test data come from the same class proportions, even if the test data does not have exactly the same feature distribution as the training data.