Approximate document outlier detection using random spectral projection

Outlier detection is an important process for text document collections, but as the collection grows, the detection process becomes a computationally expensive task. Random projection has shown to provide a good fast approximation of sparse data, such as document vectors, for outlier detection. The random samples of Fourier and cosine spectrum have shown to provide good approximations of sparse data when performing document clustering. In this article, we investigate the utility of using these random Fourier and cosine spectral projections for document outlier detection. We show that random samples of the Fourier spectrum for outlier detection provides better accuracy and requires less storage when compared with random projection. We also show that random samples of the cosine spectrum for outlier detection provides similar accuracy and computational time when compared with random projection, but requires much less storage.