Fast algorithms for topk personalized pagerank queries

Abstract

In entity-relation (ER) graphs (V, E), nodes V represent typed entities and edges E represent typed relations. For dynamic personalized PageRank queries, nodes are ranked by their steady-state probabilities obtained using the standard random surfer model. In this work, we propose a framework to answer top-k graph conductance queries. Our top-k ranking technique leads to a 4× speedup, and overall, our system executes queries 200-1600× faster than whole-graph PageRank. Some queries might contain hard predicates i.e. predicates that must be satisfied by the answer nodes. E.g., we may seek authoritative papers on public key cryptography, but only those written during 1997. We extend our system to handle hard predicates. Our system achieves these substantial query speedups while consuming only 10-20% of the space taken by a regular text index.

Reducing seed noise in personalized PageRank

“…Moreover, we also show that -these can be obtained highly efficiently, if necessary, leveraging existing approximation algorithms [2,4,14,17,21,23,41] and/or parallel implementations [3,32] for computing the PPR scores, -the proposed formulations are reuse-promoting in the sense that, it is possible to divide the work relative to individual seed nodes and cache the intermediary results obtained during the computation -these cached results can then be reused for future queries sharing seed nodes, and -especially in systems with large query throughputs, it may be possible to cluster queries based on the partial overlaps between the seed sets and, thus, significantly reduce the overall robust PPR computation costs.…”

“…Alternatively, PowerIteration [27] or using iterative approximations [14,30], which explicitly simulate the dissemination of probability mass by repeatedly applying the transition process to an initial distribution π 0 until a convergence criterion is satisfied. Recent advances on PPR computation include top-k and approximate personalized PageRank algorithms [2,4,14,17,21,23,41] and parallelized implementations on MapReduce or Pregel based systems [3,32,36,38]. The FastRWR algorithm [41], for example partitions the graph into subgraphs and indexes partial intermediary solutions.…”

“…This is especially advantageous when G is large as we can leverage any of the highly effective approximation algorithms [2,4,14,17,21,23,41] or parallelized implementations [3,32] for computing these PPR scores. Most importantly, the first step of the algorithm (where we solve a linear equation independently for each seed node) can be trivially parallelized by assigning each node to a different computation unit.…”

Locality-sensitive and Re-use Promoting Personalized PageRank computations

“…Unlike existing approximate PPR algorithms (Avrachenkov et al, 1995;Bahmani et al, 2010;Csalogany et al, 2005;Chakrabarti, 2007;Fujiwara et al, 2012;Gupta et al, 2008;Tong et al, 2006b;Song et al, 2009), LR-PPR is location sensitive. Therefore, given the set, S, of seed nodes and the corresponding localities, G S , the computation focuses on the combined locality G + (V + , E + ) ⊆ G, where…”

“…Unfortunately, for large data sets, both of these processes are prohibitively expensive. Recent advances on personalized PageRank include top-k and approximate personalized PageRank algorithms (Avrachenkov et al, 1995;Bahmani et al, 2010;Csalogany et al, 2005;Chakrabarti, 2007;Fujiwara et al, 2012;Gupta et al, 2008;Tong et al, 2006b;Song et al, 2009), parallelized implementations on MapReduce or Pregel based batch data processing systems (Bahmani et al, 2011;Malewicz et al, 2010), and techniques to eliminate seed noise (Huang et al, 2014). The FastRWR algorithm, presented in (Tong et al, 2006b), for example partitions the graph into subgraphs and indexes partial intermediary solutions.…”