| Full text | |
| Author(s): |
Mastelini, Saulo Martiello
;
Montiel, Jacob
;
Gomes, Heitor Murilo
;
Bifet, Albert
;
Pfahringer, Bernhard
;
de Carvalho, Andre C. P. L. F.
;
Xue, B
;
Pechenizkiy, M
;
Koh, YS
Total Authors: 9
|
| Document type: | Journal article |
| Source: | 21ST IEEE INTERNATIONAL CONFERENCE ON DATA MINING WORKSHOPS ICDMW 2021; v. N/A, p. 9-pg., 2021-01-01. |
| Abstract | |
Incremental Hoeffding Tree Regressors (HTR) are powerful non-linear online learning tools. However, the commonly used strategy to build such structures limits their applicability to real-time scenarios. In this paper, we expand and evaluate Quantization Observer (QO), a feature discretization-based tool to speed up incremental regression tree construction and save memory resources. We enhance the original QO proposal to create multi-branch trees when dealing with numerical attributes, creating a mix of interval and binary splits rather than binary splits only. We evaluate the multi-branch and strictly binary QO-based HTRs against other tree-building strategies in an extensive experimental setup of 15 data streams. In general, the QO-based HTRs are as accurate as traditional HTRs, incurring one-third of training time at only a fraction of the memory resource usage. The obtained numerical multi-branch HTRs are shallower than the strictly binary ones, significantly faster to train, and they keep predictive performance similar to the traditional incremental trees. (AU) | |
| FAPESP's process: | 18/07319-6 - Multi-target data stream mining |
| Grantee: | Saulo Martiello Mastelini |
| Support Opportunities: | Scholarships in Brazil - Doctorate |