OSRT: An Online Sparse Approximation Model for Scattered Data

Yufeng Zheng, Xin Xu, Xiaopeng Luo, Kanghui Zhu

Primary Area: metric learning, kernel learning, and sparse coding

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Online learning, Scattered data, Adaptive sparse approximation, Radial basis function, Tree decomposition

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Online learning is a crucial technique for dealing with large and evolving datasets in various domains, such as real-time data analysis, online advertising, or financial modeling. In this paper, we propose a novel predictive statistical model called the Online Sparse Residual Tree (OSRT) for handling streaming multivariate scattered data. OSRT is based on online tree decomposition and online adaptive radial basis function (RBF) exploration. OSRT dynamically expands its network depth as more data arrives, and incorporates a sparse and appropriate RBF refinement at each child node to minimize the residual error from its parent node. OSRT also uses an incremental method to explore the central node of the RBF function, ensuring both sparsity and accuracy of the model. When the network reaches its maximum depth, the OSRT model updates the RBF approximation of its final layer based on the most recent data. This ensures that the model captures the latest trends in the evolving data. We evaluate our algorithm on several datasets, and compare it with existing online RBF methods. From the results, it is shown that OSRT achieves higher efficiency and accuracy.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4600