| Full text | |
| Author(s): |
Total Authors: 3
|
| Affiliation: | [1] Fed Univ Sao Carlos UFSCar, Dept Comp Sci, Sorocaba, SP - Brazil
[2] Shopify Inc, Ottawa, ON - Canada
Total Affiliations: 2
|
| Document type: | Journal article |
| Source: | KNOWLEDGE-BASED SYSTEMS; v. 188, JAN 5 2020. |
| Web of Science Citations: | 0 |
| Abstract | |
In recent decades, various machine learning methods have been proposed to address classification problems. However, most of them do not support incremental (or online) learning and therefore are neither scalable nor robust to dynamic problems that change over time. In this study, a classification method was introduced based on the minimum description length principle, which offered a very good trade-off between model complexity and predictive power. The proposed method is lightweight, multiclass, and online. Moreover, despite its probabilistic nature, it can handle continuous features. Experiments conducted on real-world datasets with different characteristics demonstrated that the proposed method outperforms established online classification methods and is robust to overfitting, which is a desired characteristic for large, dynamic, and real-world classification problems. (C) 2019 Elsevier B.V. All rights reserved. (AU) | |
| FAPESP's process: | 18/02146-6 - Distributed text representation model with online learning |
| Grantee: | Renato Moraes Silva |
| Support Opportunities: | Scholarships in Brazil - Post-Doctoral |
| FAPESP's process: | 17/09387-6 - A continuously evolving distributed text representation model |
| Grantee: | Tiago Agostinho de Almeida |
| Support Opportunities: | Regular Research Grants |