无线传感器网络数据的个性化加权在线集成学习算法

Personalized Online Weighted Ensemble Learning Algorithm for Classifying Wireless Sensor Networks Data

作为一种数据采集型网络,无线传感器网络有效地提高了人们获取客观物理信息的能力,已成为智慧生活的重要一环.然而,无线传感器网络数据是一种流数据,显著特点在于具有时间顺序性,实时性高,容易受外界因素干扰,且易产生分布漂移等特点.要利用无线传感器网络建立精确可靠的预测模型,就要兼容其上述特性.为此,本文提出了一种个性化加权在线集成算法,其个性化特征在于将集成中每个个体分类器的权重与样本集分布的相似性相关联,即当接收到一个新的数据块时,首先假定其符合多维高斯分布,进而计算其与之前数个数据块的K-L散度值,也即相似度,最后根据该值来确定各个数据块所对应分类器的投票权重.通过实验分析,该方法能够在线响应无线传感器网络数据流,且与传统的在线学习算法相比,具有更高的预测精度.

As a data acquisition network,wireless sensor network has effectively improved the ability to obtain objective physical information,and has become an important part of intelligent life. However,wireless sensor network data is a kind of streaming data,which is characterized by time sequence,high real-time and distribution drift due to external factors. It is necessary to consider the characteristics above to construct an excellent predictive model for the wireless sensor data. In this article,we propose a personalized weighted online ensemble predictive algorithm. The core idea of this algorithm is to correlate the weight of each base classifier and the similarity of the instance distribution. When a new data chunk is received,we firstly assume it satisfy the multi-dimensional Gaussian distribution,further the K-L divergence,i. e.,the similarity,between it and each previous data chunk are calculated,finally according to the similarity to assign the weight for each base classifier. Based the analysis of experimental results,the algorithm can construct the better predictive model than some traditional ones.