Research On The New Algorithm Of SOM And Its Application

Self-organizing feature map is Proposed by Teuvo Kohonen Professor in the 1980s, It simulates the self-organizing feature of the brain cell to realize recognition, clustering, compositor and so on. Self-organizing map have the characteristic of Topologically Consistency and Vector Quantization. It is an effective method which describes the content of uncompact data.Twinned self-organizing maps expand traditional self-organizing maps .The paper apply algorithms which are normally only applied to a single data set to the pair of twinned sets. Then the network can perform a type of nonparametric CCA. This new network can be considered as an extended linear technology which can be applied to two related data sets. Self-organizing map has obtained many successes in practical applications, but it also has many limitations, such as, the whole training process is carried on input samples. Because algorithm is based on Euclidean distance, so the ability for classifying will be lower when the input sample has altitudinal nonlinear structure. Moreover, it can be influenced by noise and wild value . This paper focuses mainly on the reasonable design for SOM. Our work includes the following subjects:(1) Robustness about SOM. In order to enhance the robustness, euclidean distance is replaced by the distance to the Vorinoi cell in the proposed SOM . We illustrate the proposed SOM predictive power on a noisy data sets, results demonstrate the effectiveness and robustness capability of the proposed SOM.(2) Kernel methods for SOM. Kernel means, performing a nonlinear data transformation into some high dimensional feature space, increases the probability of the linear separability of the patterns within the feature space. Multiformity of kernels leads to different metrics of distance in input space, and correspondingly results in SOM classifications, Algorithm selects the different kernel functions in view of the different questions, it can extend the model. The paper illustrates the proposed SOM predictive power on a real financial time series, results demonstrate the effectiveness and robustness capability of the proposed SOM. Then the paper discusses Mahalanobis Distance Kernels for SOM.