Double Regularization-Based RVFL and edRVFL Networks for Sparse-Dataset Classification

Abstract

In our previous work, the random vector functional link network (RVFL) and the ensemble deep RVFL network (edRVFL) have been proven to be competitive for tabular-dataset classification, and their sparse pre-trained versions (SP-RVFL and SP-edRVFL) perform well for sparse-dataset (datasets with a large number of features) classification. However, the sparse auto-encoder-based versions suffer from a time-consuming problem. Therefore, we need to find an alternative way to have similar performance and faster training time. In this paper, we propose the double regularization-based RVFL (2R-RVFL) and edRVFL networks (2R-edRVFL). Two different regularization parameters are assigned to the input and hidden features, respectively. The experiments on 12 sparse datasets show that the 2R-RVFL and 2R-edRVFL networks have similar performance as the SP-RVFL and SP-edRVFL networks, and the double regularized variants have a huge training time advantage. Thus, we believe the newly proposed 2R-RVFL and 2R-edRVFL networks are more suitable for sparse dataset classification. 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.