Finding Key Knowledge Attribute Subspace of Outliers for High Dimensional Dataset

Detecting outliers is an important task in many applications. Since most applications possess high dimensional data, traditional outlier detecting methods will become inefficient in such cases. To solve the problem, we propose the concept of outlying reduction by extending attribute reduction in rough set theory. Additionally, we define the key knowledge attribute subspace (KKAS), which can produce the outlying partition approximating to that in the full dimensional attribute space. An efficient method for finding KKAS is proposed. It first finds all outliers in the full attribute space and then, calculates KAS for corresponding projection of each outlier. Finally, the KKAS can be identified by the value of outlying partition similarity. Experiments both on UCI datasets and real-life datasets show that our method is effective, efficient, and highly scalable.