存在型空值插补的特征约简方法研究

Feature reduction based on interpolated existential null values

特征约简是多源信息融合中剔除冗余数据、提高融合精度的有效途径.在对多传感器数据构成的不完备信息系统进行分析的基础上,利用多种距离分析方法计算存在型空值分布区间;针对不完备信息系统同时含有的2种空值类型问题(存在型空值和不存在型空值),提出基于存在型空值插补的限制容差关系;引入知识粒度的概念,并结合特征所包含的知识依存关系,研究基于属性重要度的启发式约简算法;通过实验验证了算法的有效性,并对其性能进行了分析.该方法同时考虑了多种空值类型,更加符合多传感器数据的特征,避免了单纯考虑不存在空值或者遗失空值造成的约简不准确问题;与其他约简算法相比,所提出的算法不仅考虑了单个的填补数值,同时还将插补数值的可能离散集合也考虑进来,增加了约简算法的适应性.

In multi-source information fusion,feature reduction is an accepted and effective way to eliminate redundant data and enhance the precision of fusion.Facing an incomplete information system containing sensor data with a multi-source heterogeneous structure,First,variety of analytical methods for calculating distances were used to determine the interval distribution of existing null values.Second,a limited tolerance relation based on interpolating existing null values was proposed for this incomplete information system.It has two kinds of null value types： an existing null value type and an unassigned value type.Next,by combining the interdependent relation of knowledge included in features,the heuristic algorithm for reducing the importance of feature attributes was improved by introducing the concept of knowledge granularity.Subsequently,performance of the algorithm was validated in an experiment.A variety of types of null values were taken into account,as they were more suitable for the features of multi-source heterogeneous sensor data.This eliminated oversimplifications of non-existent or missing null values which then cause inaccurate reductions.Compared with other reduction algorithms,the proposed heuristic algorithm based on knowledge granularity not only considers the filling of a single value,but also the possible discrete set of interpolation values.Both techniques can increase the adaptability of the reduction algorithm.