FREDE

Low-dimensional representations, or embeddings , of a graph's nodes facilitate several practical data science and engineering tasks. As such rely, explicitly implicitly, on similarity measure among nodes, they require the computation quadratic matrix, inducing tradeoff between space complexity embedding quality. To date, no graph work combines (i) linear complexity, (ii) nonlinear transform as its basis, (iii) nontrivial quality guarantees. In this paper we introduce FREDE ( FREquent Directions Embedding ), based matrix sketching that those three desiderata. Starting out from observation methods aim to preserve covariance rows iteratively improves while individually processing nonlinearly transformed PPR derived state-of-the-art method provides, at any iteration column-covariance approximation guarantees in due course almost indistinguishable optimal by SVD. Our experimental evaluation variably sized networks shows performs well SVD competitively against diverse tasks, even when it is little 10% node similarities.