面向特征演化数据流的增量学习方法研究  

Incremental Learning Method for Feature-evolving Data Streams

作　　者：陈燕菲 刘三民 CHEN Yanfei;LIU Sanmin(School of Computer and Information,Anhui Polytechnic University,Anhui Wuhu 241000,China)

机构地区：[1]安徽工程大学计算机与信息学院,安徽芜湖241000

出　　处：《重庆工商大学学报(自然科学版)》2025年第1期94-104,共11页Journal of Chongqing Technology and Business University:Natural Science Edition

基　　金：安徽省高校自然科学研究重点项目(2022AH050972,KJ2021A0516,KJ2019ZD15);安徽省自然科学基金项目(2108085MF213).

摘　　要：目的特征演化数据流的特征空间随时间推移而动态变化,传统增量学习方法囿于固定特征空间的假设,无法直接应用于特征演化数据流的学习场景,因此针对挖掘特征演化数据流时面对的分类模型与当前数据特征不匹配而失效、模型预测性能受噪声干扰等问题,提出了一种面向特征演化数据流的增量学习方法。方法首先,通过引入模糊隶度函数并结合增量孪生支持向量机模型,鲁棒地训练与更新分类器;当出现新特征时,重新训练新分类器,同时结合局部线性加权回归算法拟合新旧特征之间的映射关系,从而在旧特征消失时,利用所学到的映射关系,将已训练好的旧分类器投影至新特征空间继续更新;最后,结合两种不同的集成策略以合并新旧两分类器实现共同预测。结果通过大量仿真实验,所提方法分类准确率相较于对比方法提升了0.3%~21.7%;在含不同信噪比数据集上,分类模型性能整体优于对比模型,并随着人工增加噪声比例,模型分类效果受负面影响较小。结论所提方法得以构建性能高效稳定的分类模型,在提升模型预测精度的同时能减少噪声对分类性能的干扰,增强了模型对特征演化数据流自适应学习能力。Objective The feature space of feature-evolving data streams dynamically changes over time.Traditional incremental learning methods are constrained by the assumption of a fixed feature space and cannot be directly applied to the learning scenario of feature-evolving data streams.Therefore,to address the problems of classification models mismatching with current data features and model prediction performance being affected by noise interference when mining feature-evolving data streams,an incremental learning method tailored to feature-evolving data streams was proposed.Methods Firstly,by introducing fuzzy membership functions and combining them with an incremental twin support vector machine model,classifiers were robustly trained and updated.When new features appeared,new classifiers were retrained,and the mapping relationship between new and old features was fitted using a local linear weighted regression algorithm.Thus,when old features disappeared,the trained old classifiers were projected into the new feature space for continued updating using the learned mapping relationship.Finally,two different ensemble strategies were combined to merge the new and old classifiers for joint prediction.Results Through extensive simulation experiments,the proposed method improved classification accuracy by 0.3% to 21.7% compared with baseline methods.On datasets with different signal-to-noise ratios,the overall performance of the classification model was superior to that of the baseline model,and the model's classification effectiveness was less affected by artificially increasing the noise ratio.Conclusion The proposed method is verified to construct an efficient and stable classification model,which not only enhances model prediction accuracy but also reduces the interference of noise on classification performance,thus strengthening the adaptive learning capability of the model for feature-evolving data streams.

关 键 词：数据流挖掘 特征演化 增量学习 动态数据流 集成学习 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]