Making large-scale support vector machine learning practical
Advances in kernel methods
A patent search and classification system
Proceedings of the fourth ACM conference on Digital libraries
IEEE Transactions on Neural Networks
Hi-index | 0.00 |
With the adoption of min-max-modular support vector machines (SVMs) to solve large-scale patent classification problems, a novel, simple method for incorporating prior knowledge into task decomposition is proposed and investigated. Two kinds of prior knowledge described in patent texts are considered: time information, and hierarchical structure information. Through experiments using the NTCIR-5 Japanese patent database, patents are found to have time-varying features that considerably affect classification. The experimental results demonstrate that applying min-max modular SVMs with the proposed method gives performance superior to that of conventional SVMs in terms of training time, generalization accuracy, and scalability.