多变量时间序列的无监督属性选择算法

摘要

属性选择是一种有效的数据预处理方法，可同时保留多变量时间序列重要变量的时序关系及其实际物理意义。针对很多实际数据无类别信息的问题，文中提出一种无监督属性选择算法并分析其复杂度。首先设计一种无需进行相空间重构的多变量时间序列分形维数计算方法，并将分形维数视为其本质维，利用属性子集的分形维数及其属性数目的变化作为子集优劣的评价标准。再优化离散粒子群算法以解决高维属性空间搜索的“组合爆炸”问题。最后利用典型混沌动力学系统所产生的多变量时间序列和UCI数据库的5组数据集进行仿真计算，结果表明该算法可在较短时间内找到较优的属性子集，具有较优的整体性能。%Attribute selection is an effective data preprocessing method. It can keep temporal relations of important attributes of multivariate time series and their actual physical meanings. Aiming at the problem that the actual data lacks the classified information, an unsupervised attribute selection method is proposed and its time complexity is analyzed. Firstly, a method for computing the fractal dimension of multivariate time series is proposed, and there is no need for the proposed method to reconstruct the phase space. The fractal dimension is considered as the essential dimension by the proposed method. Therefore ,the changing of the attributes number and the fractal dimension of attribute subsets are regarded as the evaluation criterion of attribute subsets. To solve the combinatorial explosion problem in high dimensional search space, the discrete particle swarm optimization algorithm is improved. Finally, the results of numerical simulations of multivariate time series from the typical chaotic dynamic system and five datasets of UCI database confirm the effectiveness of the proposed algorithm. Moreover, experimental results show the proposed algorithm finds out better attributes sets in shorter time and achieves better integrative performance.