Non linear dimensional reduction method based on supervised neighborhood graph

In this paper, we proposed a novel supervised method to construct ‘neighborhood graph’, which is often constructed in recent non-linear dimensional reduction techniques. The key ideas in our proposed method is introducing a new distance criterion based on weighted Euclidean distance between data points, which use class label information of data points. In order to evaluate, the proposed method was used as the primary stages of non-linear dimensional reduction techniques in LLE and Isomap. The proposed method was tested on four artificial data sets which are conventional in dimensional reduction research and the results were compared with the results of some unsupervised linear and non-linear dimensional techniques and some supervised linear dimensional reduction techniques. Although the tests are performed on artificial data sets in this paper, the proposed method could be applied to other problems such as face recognition and body pose for instance. Results of experiments illustrated that using the neighborhood graph obtained our proposed method improves the results of existing non-linear dimensional reduction techniques.