garfield-kwok / fora

FORA: Simple and Effective Approximate Single-Source Personalized Pagerank

Code Contributors: Sibo Wang, Renchi Yang

If you have any questions, feel free to contact us. Our emails are: {wangsibo.victor, anryyang}@gmail.com.

Plase cite our papers if you choose to use our code.

@inproceedings{WANGYXWY17,
  author    = {Sibo Wang and
               Renchi Yang and
               Xiaokui Xiao and
               Zhewei Wei and
               Yin Yang},
  title     = {{FORA:} Simple and Effective Approximate Single-Source Personalized
               PageRank},
  booktitle = {{SIGKDD} 2017},
  pages     = {505--514},
  year      = {2017},
}

@article{WANGYWXWLY019,
  author    = {Sibo Wang and
               Renchi Yang and
               Runhui Wang and
               Xiaokui Xiao and
               Zhewei Wei and
               Wenqing Lin and
               Yin Yang and
               Nan Tang},
  title     = {Efficient Algorithms for Approximate Single-Source Personalized PageRank
               Queries},
  journal   = {{ACM} Trans. Database Syst.},
  volume    = {44},
  number    = {4},
  pages     = {18:1--18:37},
  year      = {2019},
}

Single Source Personalized PageRank

Tested Environment

• Ubuntu
• C++ 11
• GCC 4.8
• Boost (Boost 1.65 on our machine)
• cmake

Compile

$ cmake .
$ make

Parameters

./fora action_name --algo <algorithm> [options]

• action:
  • query
  • topk
  • build: build index, for FORA only
  • generate-ss-query: generate queries file
  • gen-exact-topk: generate ground truth by power-iterations method
  • batch-topk: for different k=i*config.k/5, i=1,2,3,4,5, compute precision
• algo: which algorithm you prefer to run
  • bippr: Bidirectional PPR
  • fora: FORA
  • montecarlo: Monte Carlo
  • fwdpush: Forward Push
  • hubppr: HubPPR
• options
  • --prefix <prefix>
  • --epsilon <epsilon>
  • --dataset <dataset>
  • --query_size <queries count>
  • --k <top k>
  • --with_idx: for FORA or HubPPR
  • --hub_space <hubppr oracle space-consumption>
  • --exact_ppr_path <directory to place generated ground truth>
  • --result_dir <directory to place results>
  • --balanced: a balance strategy is used to automatically decide R_max for FORA.
  • --opt: optimization techniques for whole-graph SSPPR and top-k queries are applied.

Data

The example data format is in ./data/webstanford/ folder. The data for DBLP, Pokec, Livejournal, Twitter are not included here for size limitation reason. You can find them online.

Generate queries

Generate query files for the graph data. Each line contains a node id.

$ ./fora generate-ss-query --prefix <data-folder> --dataset <graph-name> --query_size <query count>

• Example:

$ ./fora generate-ss-query --prefix ./data/ --dataset webstanford --query_size 1000

Indexing

Construct index files for the graph data using a single core. (Only for FORA)

$ ./fora build --prefix <data-folder> --dataset <graph-name> --epsilon <relative error> (--opt)

Note: the code of indexing for hubppr is not included here, please turn to the code of hubppr: https://sourceforge.net/projects/hubppr/

• Example: build index for KDD version.

$ ./fora build --prefix ./data/ --dataset webstanford --epsilon 0.5

• Example: build index for TODS version.

$ ./fora build --prefix ./data/ --dataset webstanford --epsilon 0.5 --opt

Query

Process queries.

$ ./fora query --algo <algo-name> --prefix <data-folder> --dataset <graph-name> --result_dir <output-folder> --epsilon <relative error> --query_size <query count> (--opt)

• Example:

// without index KDD version
$ ./fora query --algo fora --prefix ./data/ --dataset webstanford --epsilon 0.5 --query_size 20

// without index TODS version (balance strategy and optimization technique included)
$ ./fora query --algo fora --prefix ./data/ --dataset webstanford --epsilon 0.5 --query_size 20 --balanced --opt

// with index KDD version
$ ./fora query --algo fora --prefix ./data/ --dataset webstanford --epsilon 0.5 --query_size 20 --with_idx

// with index TODS version
$ ./fora query --algo fora --prefix ./data/ --dataset webstanford --epsilon 0.5 --query_size 20 --with_idx --opt

Top-K

Process top-k queries.

$ ./fora topk --algo <algo-name> --prefix <data-folder> --dataset <graph-name> --result_dir <output-folder> --epsilon <relative error> --query_size <query count> --k <k>

• Example

// without index
$ ./fora topk --algo fora --prefix ./data/ --dataset webstanford --epsilon 0.5 --query_size 20 --k 500

// without index TODS version
$ ./fora topk --algo fora --prefix ./data/ --dataset webstanford --epsilon 0.5 --query_size 20 --k 500 --opt

// with index KDD version
$ ./fora topk --algo fora --prefix ./data/ --dataset webstanford --epsilon 0.5 --query_size 20 --k 500 --with_idx

// with index TODS version
$ ./fora topk --algo fora --prefix ./data/ --dataset webstanford --epsilon 0.5 --query_size 20 --k 500 --with_idx --opt

Exact PPR (ground truth)

Construct ground truth for the generated queries.

$ ./fora gen-exact-topk --prefix <data-folder> --dataset <graph-name> --k <k> --query_size <query count> --exact_ppr_path <folder to save exact ppr>

• Example

$ mkdir ./exact
$ ./fora gen-exact-topk --prefix ./data/ --dataset webstanford --k 1000 --query_size 100 --exact_ppr_path ./exact/

Batch Top-K

Specify k value, and process queries for various k values (k/5, 2k/5, 3k/5, 4k/5, k), the folder contains ground truth need to be specified to obtain precision.

$ ./fora batch-topk --algo <algo-name> --prefix <data-folder> --dataset <graph-name> --result_dir <output-folder> --epsilon <relative error> --query_size <query count> --k <k> --exact_ppr_path <folder contains ground truth>

• Example

$ ./fora batch-topk --algo fora --prefix ./data/ --dataset webstanford --epsilon 0.5 --query_size 20 --k 500 --exact_ppr_path ./exact/