Towards expanding relevance vector machines to large scale datasets.

Area of Science:

• Machine Learning
• Natural Language Processing
• Computational Statistics

Background:

• Relevance Vector Machines (RVM) offer state-of-the-art performance with sparse, probabilistic solutions.
• Previous applications of RVM to large-scale text data faced computational challenges.
• Bayesian inference underpins RVM learning, enabling efficient model representations.

Purpose of the Study:

• To develop and analyze methods for scaling automated learning of RVM to large text datasets.
• To overcome computational constraints that have limited RVM application in big data scenarios.
• To enhance classification performance by leveraging extensive training data.

Main Methods:

• Proposed a diversified set of divide-and-conquer strategies for RVM.
• Employed decomposition techniques to create smaller, manageable working sets.
• Explored incremental, ensemble, and boosting approaches for improved learning.

Main Results:

• Demonstrated performance gains on benchmark datasets (Reuters-21578, RCV1).
• Maintained sparse solutions characteristic of RVM.
• Validated the effectiveness of proposed methods for large-scale text classification.

Conclusions:

• The developed methods successfully scale RVM to large text datasets.
• Divide-and-conquer strategies combined with ensemble methods enhance classification.
• The sparse and efficient solutions are suitable for deployment in distributed environments.