A Multi-grained Cascade Structure for Online Sparse Streaming Feature Selection

Online Streaming Feature Selection (OSFS) is an efficient way to handle high-dimensional streaming data. While in real applications, streaming data often have a large amount of missing data due to various reasons, leading to the accuracy loss of OSFS. Latent Factor Analysis (LFA) is one of the most effective methods to pre-estimate the missing data of sparse streaming data. However, the known data density of sparse streaming features is always low, which is not enough for LFA to achieve the satisfactory pre-estimating performance for OSFS. Motivated by this, this paper proposes a sparse OSFS algorithm based on a multi-grained cascaded structure, termed as SO-MCS. Its main idea is two-fold: 1) using LFA to pre-estimate the missing data in sparse flow features in advance, and 2) gradually generating synthetic data to construct a multi-layer LFA architecture, which develops the known data density of the input sparse streaming feature matrix. Extensive experiments are carried out on eight real-world datasets, demonstrating that SO-MCS outperforms five state-of-the-art OSFS algorithms to a significant extent.

I. Introduction

With the development of advanced technology, the handling of high-dimensional data faces significant challenges [1] - [7]. Feature selection can remove irrelevant and redundant features from high-dimensional data to reduce time and space complexity. Traditional feature selection algorithms assume a known feature space dimension [8]. However, in practical data processing, features flow one after another over time. To address this issue, an online streaming feature selection (OSFS) approach has been proposed [9].