Learning framework based on ER Rule for data streams with generalized feature spaces

• Published in: Information sciences 2023-11, Vol.649, p.119604, Article 119604
• Main Authors: Zhao, Rui-Rui, Sun, Jian-Bin, You, Ya-Qian, Jiang, Jiang, Yu, Hai-Yue
• Format: Article
• Language: English
• Subjects: Compact representation; Data stream learning; Evidential reasoning rule; Feature selection; Generalized feature spaces
• ISSN: 0020-0255; 1872-6291
• DOI: 10.1016/j.ins.2023.119604

Learning with data streams has recently been the focus of extensive research and various solutions have been proposed. However, most such studies assume that the features remain unchanged or change with stringent hypotheses, which cannot always hold in practical applications. Considering the proven superiority of Evidential Reasoning Rule (ER Rule) classifier and its parameters are separately associated with each feature, a learning framework based on the ER Rule for data streams with Generalized Feature Spaces (ER Rule-GFS) is developed. In ER Rule-GFS, the feature spaces of data streams are generalized, which means that at each stage, new features and existing features will occur randomly or not. When a new data stream arrives, a certain effective feature selection algorithm is used to select the informative features first. Then, the selected features and the features stored thus far are divided into three categories, namely, active/new/silent. According to whether active and new are null, three generalized basic learning problems are formulated, and their solutions are given. Afterwards, if there are too many features, follow-on feature selection will be performed mainly in silent based on the activated frequency and weight. The effectiveness of ER Rule-GFS is verified by various experiments on 20 synthetic datasets.