| Description |
Matrices are essential data representations for many large-scale problems in data analytics; for example, in text analysis under the bag-of-words model, a large corpus of documents are often represented as a matrix. Many data analytic tasks rely on obtaining a summary (a.k.a sketch) of the data matrix. Using this summary in place of the original data matrix saves on space usage and run-time of machine learning algorithms. Therefore, sketching a matrix is often a necessary first step in data reduction, and sometimes has direct relationships to core techniques including PCA, LDA, and clustering. In this dissertation, we study the problem of matrix sketching over data streams. We first describe a deterministic matrix sketching algorithm called FrequentDirections. The algorithm is presented an arbitrary input matrix A∈ Rn&× d one row at a time. It performs O(dl) operations per row and maintains a sketch matrix B ∈ Rl× d such that for any k< l, ||ATA - BTB \|| 2 < ||A - Ak||F2 / (l-k) and ||A - πBk(A)||F2 ≤ (1 + k/l-k)||A-Ak||F2 . Here, Ak stands for the minimizer of ||A - Ak||F over all rank k matrices (similarly Bk), and πBk (A) is the rank k matrix resulting from projecting A on the row span of Bk. We show both of these bounds are the best possible for the space allowed, the sketch is mergeable, and hence trivially parallelizable. We propose several variants of FrequentDirections that improve its error-size tradeoff, and nearly matches the simple heuristic Iterative SVD method in practice. We then describe SparseFrequentDirections for sketching sparse matrices. It resembles the original algorithm in many ways including having the same optimal asymptotic guarantees with respect to the space-accuracy tradeoff in the streaming setting, but unlike FrequentDirections which runs in O(ndl) time, SparseFrequentDirections runs in Õ(nnz(A)l + nl2) time. We then extend our methods to distributed streaming model, where there are m distributed sites each observing a distinct stream of data, and which has a communication channel with a coordinator. The goal is to track an ε-approximation (for ε ∈ (0,1)) to the norm of the matrix along any direction. We present novel algorithms to address this problem. All our methods satisfy an additive error bound that for any unit vector x, | ||A x||2 - ||B x ||2 | ≤ |ε ||A||F2 holds. |
