Reinforced quasi-random forest. (October 2019)

• Record Type: Journal Article
• Authors: Paul, Angshuman; Mukherjee, Dipti Prasad
• Is Part Of: Pattern recognition. Volume 94(2019:Oct.)
• Page Start: 13
• Page End: 24
• Publication Date: 2019-10
• Subjects: Random forest -- Reinforcement learning -- Orthogonal decision trees -- Importance of attributes; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2019.05.013
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A reinforcement strategy for random forest. A novel bottom-up approach to grow a set of mutually orthogonal trees. Introduction of quasi-random trees. An electrostatic model to find trees for reinforcement. Abstract: We propose a reinforced quasi-random forest for classification task. Reinforcement is performed iteratively by adding new trees to the forest. Our method assigns an importance to each of the attributes and identifies the attributes that causes the mis-classification of data points during training. The new trees are constructed using the mis-classified data points with reduced set of attributes. The attributes for splitting the nodes of the reinforced trees are found in a deterministic manner. Hence the new trees are quasi-random in nature. The best out of all the new trees are found using a novel electrostatic model. These trees are termed as reinforced trees. Additions of reinforced trees to the existing forest ensure maximum reduction in classification error. The efficacy of the proposed method is established through experiments on breast cancer datasets for detecting mitotic nuclei. Results of our method show significant improvement compared to other state-of-the-art approaches. Results on benchmark datasets show as much as 14% reduction in classification error.