Feature matricization for document classification

Generally, the dimension of feature vector in text classification depends on the number of words in the specific domain. Many documents of considered categories make it numerous. Therefore, the dimension of feature vector is very high that makes it consumes a lot of time and memory to process. Moreover, it is a cause of the small sample size problem when the number of available training documents is far smaller than the dimension of these feature vectors. This paper proposes the alternative technique of dimensionality reduction for the feature vector in two-dimensional manner by previously transforming the feature vector to the feature matrix and then using Two-Dimensional Principal Component Analysis (2DPCA) for reducing the dimension of this feature matrix. Based on 2DPCA, the original weighted term matrix is not necessary to store in the memory anymore because the scatter matrix of 2DPCA can be computed incrementally. The small reduction in matrix form impacts to the plenty of dimensionality reduction in vector form. From the experimental results on well-known dataset, the proposed method not only significantly reduce the dimensionality but also achieve the higher accuracy rate than the original feature space.