Subspace Learning Algorithms Based On Self-organizing MAP

With the increasingly growth of dimensionality,“the curse of dimensionality” seems an inevitable problem for many data mining tasks.Dimensionality reduction draws much attention from researchers.Instead of original high-dimensional input space,selecting partial dimensions helps to reduce the computation cost and improve the performance of models.Feature selection is an important subspace method that could remove dimensions with no relevance or redundancy,but in the risk of information loss.Also,it cannot deal with problem that some dimensions only work on partial data and turn into noise for other.In order to discover and apply the hidden subspace structure for better use,other subspace methods are expected for dimensionality reduction.First,this paper gives an overview of subspace learning methods and a detailed description of Self-Organizing Map(SOM)neural network.Then a subspace clustering method based on SOM neural network is proposed,which searches on the SOM map directly.It first seeks for sub regions in each dimension and then merges them iteratively.Experimental results show that this algorithm is able to find subspace clusters effectively with higher accuracy compared with original SOM neural network and several other subspace clustering methods.Besides,it also shows good adaptability on different datasets.Next,a subspace ensemble classification method based on SOM neural network is proposed,which focuses on the relationships among dimensions.On this basis,data space is divided into subspaces.The final classification result is an integration of output from each subspace.Subspaces differ with an independent classifier for each,which guarantees the diversity of the ensemble.Classifiers are trained in subspace to alleviate high dimensional pressure and reduce computational cost.According to experiments,this algorithm outperforms other classification methods and selects smaller subspace compared with other subspace algorithms.Based on the output of SOM neural network,this paper focuses on data structure and distribution in subspace to reduce interference from noise and discrete point data.Through dimension reduction and precise feature subspace for data points,algorithms for classification and clustering will achieve higher performance.