Online learning for search and classification
Thành Tâm Nguyên · 2013
Online learning is a common and useful tool for machine learning and data mining.In contrast to batch learning, online learning receives a sequence of training instances and uses some of them at a time.By the nature of online learning, the training instances may be processed only once.Therefore online learning algorithms can work on big data beyond the memory or disk capacity as well as streaming data.Moreover in document classification, online linear learning has been shown to be much more efficient than non-linear learning in terms of training and testing time.Therefore, online linear learning has recently become an active research topic.This thesis proposes a research framework that attempts to solve the search and classification problems based on the online linear learning approaches.Specifically, we have proposed online learning classification algorithms that are able to work on multiple view datasets and an online learning-to-rank algorithm that improves the accuracy of a search engine.The main research contributions are listed as follows:• Feature selection.We have investigated a number of newly supervised term weighting methods to improve the performance of text classification.These methods are evaluated on a number of text datasets and compared with other well-known unsupervised and supervised term weighting methods including tf × idf and tf × rf .• Online classification.We have proposed several online learning algorithms that can be used for topic classification.The proposed online algorithms were shown to outperform existing online learning algorithms on benchmarked datasets such as letter recognition and sentiment analysis.• Two-view online learning.We have proposed a two-view online learning algorithm, which can work on two-view datasets.The algorithm was evaluated on multiple view datasets such as Web page classification and math topic classification.The experimental results shown that it worked better than other competitors did.• Online learning-to-rank.For search engine, we have proposed an online learning-torank algorithm, which was to learn a scoring function to re-rank the search result.The proposed algorithm improved the accuracy of a search system on a math document dataset.In summary, our proposed approaches have been benchmarked against competing algorithms, outperforming them on numerous real-life datasets.However, these are only preliminary successes.For future work, we will continue with our investigation on feature selection, online classification, multiple view online learning, and online learning-to-rank.i