Abstract:Aiming at the problem of high probability of negative impact about redundant attributes in random forest algorithms, a Three-way Selection Random Forest algorithm based on decision boundary entropy (TSRF) is proposed without losing randomness and reducing the influence of redundant attributes on decision-making results. According to the characteristics of the attribute, the concept of decision boundary entropy is defined. Then a measuring method of attribute importance based on decision boundary entropy is proposed and set as an evaluation standard. Three-way decision is constructed and the attribute is divided into three candidate domains, namely positive domain, negative domain and boundary domain. In order to ensure the randomness of attributes, three-way attribute random selection rules based on attribute randomness are established and a certain number of attributes are randomly selected from the three candidate domains. Combine the samples selected by the bootstrap sampling method with attribute sets selected by three-way decision to produce training sample sets so that we can train the decision trees and generate forest. Six datasets are selected for the experiment. Two parameters of attribute randomness and three-way decision thresholds are analyzed to verify the theoretical conclusions respectively. The results show that the TSRF algorithm can meet the different requirements of different data sets by adjusting the parameters. The classification effect on the binary data is basically the same as the comparison algorithm, but TSRF has a significant improvement effect on the multi-class data compared with other algorithms. The proposed TSRF algorithm widens the idea for the measurement method of significance of attribute, innovates the random forest three-way selection integration method, and provides a better model framework for solving multi-classification problems.

A New Random Forest Ensemble of Intuitionistic Fuzzy Decision Trees

A New Rotation Forest Ensemble Algorithm

MTForest: Ensemble Decision Trees based on Multi-Task Learning

Dynamic Ensemble Selection Based Improved Random Forests for Fault Classification in Industrial Processes

A new random forest method based on belief decision trees and its application in intention estimation

Random Ensemble of Fuzzy Rule-Based Models.

Incremental Fuzzy Cluster Ensemble Learning Based on Rough Set Theory.

Deep ensemble forests for industrial fault classification

2 Three Techniques to Construct Rough Decision Forests

Towards generating random forests via extremely randomized trees

Induction of a Novel Hybrid Decision Forest Model Based on Information Theory.

Ensemble of optimal trees, random forest and random projection ensemble classification

Feature-Weighting and Clustering Random Forest

Constructing rough decision forests

Three-way Selection Random Forest Algorithm Based on Decision Boundary Entropy

A Survey of Fuzzy Decision Tree Classifier Methodology

Ois-Rf: A Novel Overlap And Imbalance Sensitive Random Forest

Three-Branch Random Forest Intrusion Detection Model

Ensemble of Decision Trees Based on Representative Data

Fusing multiple interval-valued fuzzy monotonic decision trees

Quasi-Adaptive Random Forest for Classification