Indonesia Conference Directory

In the supervised machine learning framework sufficient labeled data are needed to achieve satisfying performance. In many domains, however, the labeled data are often expensive to obtain and requires laborious human effort. In this paper, we propose a query technique by introduce the expected log-likelihood as a criterion to minimize the training data in active machine learning framework. In this scenario, the expected log-likelihood is calculated for all unlabeled data points then an instant with highest value is chosen to be labeled. Some experiments were conducted to test the effectiveness of the method using the Gaussian mixture model classifiers with different number of kernels for each class. The method was applied on a training and test data set generated from the two normal distributions with the same proportion for each class. The result of the experiments using the proposed method show a significant reducing of the labeled data compared with the passive learning method. On the data set, the passive learning reached an error rate of 4.50% after 81 queries, meanwhile the same error was reached after 36 queries by the active learning technique or 1,25 times faster than passive learning.