AutoSplit: Fast and Scalable Discovery of Hidden Variables in Stream and Multimedia Databases

For discovering hidden (latent) variables in real-world, nongaussian data streams or an n-dimensional cloud of data points, SVD suffers from its orthogonality constraint. Our proposed method, “AutoSplit”, finds features which are mutually independent and is able to discover non-orthogonal features. Thus, (a) finds more meaningful hidden variables and features, (b) it can easily lead to clustering and segmentation, (c) it surprisingly scales linearly with the database size and (d) it can also operate in on-line, single-pass mode. We also propose “Clustering-AutoSplit”, which extends the feature discovery to multiple feature/bases sets, and leads to clean clustering. Experiments on multiple, real-world data sets show that our method meets all the properties above, outperforming the state-of-the-art SVD.