A Fast Relative Density Method Based on Space Partitioning

Abstract

Label noise play an important role in classification. It can cause overfitting of learning methods and deteriorate their generalizability. The relative density method is effective in label noise detection, but it has high time complexity. On the other hand, the multi-granularity relative density method reduces the time cost, but the classification accuracy is also reduced. In this paper, we propose an improved relative density method, named the relative density method based on space partitioning (SPRD). The proposed method not only accelerates the label noise detection but also maintains a good classification performance. Also, the parameter k, which is used in the conventional relative density methods, is removed, making the proposed method adaptive. The experiment results on the UCI datasets demonstrate that the proposed method has higher efficiency than the conventional methods and better classification accuracy than the multi-granularity relative density method.