There is no rational distance metric for nominal variables in traditional kNN classification algorithm. And the weighting methods commonly used in kNN cannot process datasets with multi-type variables or depend much on the field knowledge. An improved kNN method based on conditional probability and QPSO is presented is this paper. This approach measures the distance between two nominal variables by the distribution difference of instances' classes, which makes full use of attribute values'

more »

... mation. Meanwhile, it adopts QPSO to adjust attribute weight so that the weight will enhance the classification accuracy. This approach is able to process datasets with multi-type variables and less depends on parameters. Finally, experiments were taken on the UCI data set, which shows that our approach is superior in performance to algorithms compared. A is the maximum and minimum