aPPR helps you calculate approximate personalized pageranks from large graphs, including those that can only be queried via an API. aPPR additionally performs degree correction and regularization, allowing users to recover blocks from stochastic blockmodels.


You can install the development version from GitHub with:

Find the personalized pagerank of a node in an igraph graph

This is a basic example which shows you how to solve a common problem:

Why should I use aPPR?

  • curious about nodes important to the community around a particular user who you wouldn’t find without algorithmic help

  • 1 hop network is too small, 2-3 hop networks are too large (recall diameter of twitter graph is 3.7!!!)

  • want to study a particular community but don’t know exactly which accounts to investigate, but you do have a good idea of one or two important accounts in that community

aPPR calculates an approximation

comment on p = 0 versus p != 0

Find the personalized pagerank of a Twitter user using rtweet

appr_rtweet(graph, "fchen365", epsilon = 1e-4, verbose = TRUE)

Advice on choosing epsilon

Number of unique visits as a function of epsilon, wait times, runtime proportion to 1 / (alpha * epsilon), etc, etc

speaking strictly in terms of the p != 0 nodes

1e-4 and 1e-5: finishes quickly, neighbors with high degree get visited 1e-6: visits most of 1-hop neighborhood. finishes in several hours for accounts who follow thousands of people with ~10 tokens. 1e-7: visits beyond the 1-hop neighbor by ???. takes a couple days to run with ~10 tokens. 1e-8: visits a lot beyond the 1-hop neighbor, presumably the important people in the 2-hop neighbor, ???

the most disparate a users interests, and the less connected their neighborhood, the longer it will take to run aPPR


  • Connected graph assumption, what results look like when we violate this assumption
  • Sampling is one node at a time

Speed ideas

compute is not an issue relative to actually getting data

Compute time ~ access from Ram time << access from disk time << access from network time.

Make requests to API in bulk, memoize everything, cache / write to disk in a separate process?

General pattern: cache on disk, and also in RAM

Working with Tracker objects


Ethical considerations

people have a right to choose how public / visible / discoverable their information is. if you come across interesting users who are not in the public eye, do not elevate them into the public eye or increase attention on their accounts without their permission.


  1. Andersen, R., Chung, F. & Lang, K. Local Graph Partitioning using PageRank Vectors. 2006. pdf

  2. Chen, F., Zhang, Y. & Rohe, K. Targeted sampling from massive Blockmodel graphs with personalized PageRank. 2019. pdf