[Paper Review] Optimized on-line computation of PageRank algorithm

Dohy Hong|arXiv (Cornell University)|Feb 28, 2012

Complex Network Analysis Techniques21 references16 citations

TL;DR

This paper proposes a novel fluid diffusion-based algorithm for accelerating PageRank computation by reordering matrix-vector updates to prioritize nodes with high fluid flow, achieving significant speedups over traditional methods like power iteration and OPIC. The method enables asynchronous distributed computation and reduces convergence time by up to 50% on real-world web graphs, especially in high-dangling-node scenarios.

ABSTRACT

In this paper we present new ideas to accelerate the computation of the eigenvector of the transition matrix associated to the PageRank algorithm. New ideas are based on the decomposition of the matrix-vector product that can be seen as a fluid diffusion model, associated to new algebraic equations. We show through experiments on synthetic data and on real data-sets how much this approach can improve the computation efficiency.

Motivation & Objective

• To address the high computational cost of computing the PageRank eigenvector in large-scale web graphs.
• To improve convergence speed of iterative PageRank algorithms beyond standard power iteration and OPIC methods.
• To design a method that is independent of update order and naturally supports asynchronous distributed computation.
• To reduce reliance on matrix completion and vector renormalization, which increase computational overhead.
• To validate the method's efficiency on synthetic and real-world datasets, including web graphs with varying link density and dangling node ratios.

Proposed method

• The method models PageRank computation as a fluid diffusion process, where fluid is propagated from nodes with high fluid mass to their outgoing links.
• It uses a fluid vector $ F_n $ and a history vector $ H_n $ to track accumulated contributions, with updates based on $ F_n = (I - J_{i_n} + dP J_{i_n}) F_{n-1} $, where $ i_n $ is the selected node for update.
• The algorithm selects nodes for update based on fluid mass, using strategies like argmax ($ \text{ALGO-MAX} $), random selection ($ \text{ALGO-RAND} $), or a heuristic combining fluid mass and link degree ($ \text{ALGO-OP} $).
• The fluid update mechanism ensures that each update uses the latest partial results, similar to Gauss-Seidel, but in a diffusion-style order that avoids dependency on update sequence.
• The method avoids matrix completion for dangling nodes by directly handling zero-column states in the transition matrix $ P $, reducing preprocessing cost.
• The algorithm is designed to be naturally deployed in asynchronous distributed systems, as update order does not affect convergence or correctness.

Experimental results

Research questions

• RQ1 Can a fluid diffusion model of matrix-vector multiplication outperform standard power iteration in PageRank computation?
• RQ2 Does an order-independent, asynchronous update strategy based on fluid mass lead to faster convergence than existing online methods?
• RQ3 How does the method perform on graphs with high proportions of dangling nodes, where traditional methods suffer from inefficiency?
• RQ4 Can heuristic node selection strategies like $ \text{ALGO-OP} $, which consider both fluid mass and link degree, further accelerate convergence?
• RQ5 What is the impact of fluid mass-based node prioritization on convergence speed compared to random or uniform update sequences?

Key findings

• On synthetic datasets with high dangling node ratios (e.g., S1, S2, S3), the proposed method achieved up to 50% reduction in elementary steps compared to OPIC and power iteration.
• On the gr0.California web graph (9,664 nodes, 16,150 links), ALGO-OP outperformed OPIC by eliminating the $ 1/\sqrt{n} $ convergence fluctuation, showing faster and more stable convergence.
• On the larger uk-2007-05@1000000 dataset (1M nodes, 41M links), ALGO-MAX outperformed ALGO-OP, indicating that optimal node selection strategy depends on graph structure and density.
• The method achieved significant speedups without requiring matrix completion or vector renormalization, reducing preprocessing and iteration costs.
• The fluid diffusion model enables natural deployment in asynchronous distributed systems, as update order does not affect convergence or correctness.
• Replacing argmax with a threshold-based batch update strategy maintained similar performance gains while reducing computational cost of node selection.