Liu Yanfang, Li Wenbin, Gao Yang. Passive-Aggressive Learning with Feature Evolvable Streams[J]. Journal of Computer Research and Development, 2021, 58(8): 1575-1585. DOI: 10.7544/issn1000-1239.2021.20210330
Citation:
Liu Yanfang, Li Wenbin, Gao Yang. Passive-Aggressive Learning with Feature Evolvable Streams[J]. Journal of Computer Research and Development, 2021, 58(8): 1575-1585. DOI: 10.7544/issn1000-1239.2021.20210330
Liu Yanfang, Li Wenbin, Gao Yang. Passive-Aggressive Learning with Feature Evolvable Streams[J]. Journal of Computer Research and Development, 2021, 58(8): 1575-1585. DOI: 10.7544/issn1000-1239.2021.20210330
Citation:
Liu Yanfang, Li Wenbin, Gao Yang. Passive-Aggressive Learning with Feature Evolvable Streams[J]. Journal of Computer Research and Development, 2021, 58(8): 1575-1585. DOI: 10.7544/issn1000-1239.2021.20210330
2(College of Mathematics and Information Engineering, Longyan University, Longyan, Fujian 364012)
Funds: This work was supported by the National Key Research and Development Program of China (2018AAA0100905), the Education Scientific Research Project of Young Teachers of Fujian Province (JAT190743), and the Science and Technology Project of Longyan City (2019LYF13002, 2019LYF12010).
In many real-world applications, data are collected in the form of a feature evolvable stream. For instance, old features of data gathered by limited-lifespan sensors disappear and new features emerge at the same time along with the sensors exchanging simultaneously. Online passive-aggressive algorithms have proven to be effective in learning linear classifiers from datasets with both a fixed feature space and a trapezoidal feature space. Therefore, in this paper we propose a new feature evolvable learning based on passive-aggressive update strategy (PAFE), which utilizes the margin to modify the current classifier. The proposed algorithm learns two models through passive-aggressive update strategy from the current features and recovered features of the vanished features. Specifically, we both recover the vanished features and mine the initialization of the current model from the overlapping periods in which both old and new features are available. Furthermore, we use two ensemble methods to improve performance: combining the predictions from the two models, and dynamically selecting the best single prediction. Experiments on both synthetic and real data validate the effectiveness of our proposed algorithm.