Title: An improved deep forest classification algorithm

Authors: Jiaman Ding; Cuihua Liu; Runxin Li; Jinguo You; Lianyin Jia

Addresses: Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming, 650500, Yunnan, China ' Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming, 650500, Yunnan, China ' Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming, 650500, Yunnan, China ' Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming, 650500, Yunnan, China ' Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming, 650500, Yunnan, China

Abstract: Recent studies suggest that deep forest is also a kind of deep learning while overcoming the problems of excessive parameters and hard-to-adjust parameters in deep neural networks. However, the deep forest does not consider the classification contributions of each forest or the time and memory consumption caused by high-dimensional features, thus leading to inefficiency. To solve these problems, an improved deep forest (IgcForest) classification algorithm is proposed in this paper. IgcForest first preserves the class distribution vectors of the main features by a pooling strategy to realise the feature reduction and reuse. Next, an adaptive weighting strategy is proposed to calculate the weight of each forest in the cascade structure. Experiments are carried out on multiple UCI datasets to verify the effectiveness of IgcForest and the results demonstrate that IgcForest achieves better results in different evaluation indexes than other algorithms.

Keywords: deep learning; deep forest; ensemble methods; pooling strategy; self-adaptive differential evolution algorithm.

DOI: 10.1504/IJMIC.2022.125548

International Journal of Modelling, Identification and Control, 2022 Vol.40 No.4, pp.305 - 314

Received: 12 Jul 2021
Accepted: 27 Oct 2021
Published online: 14 Sep 2022 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article