Edge-Weighted Personalized PageRank

This release contains implementations from paper "Edge-Weighted Personalized PageRank: Breaking A Decade-Old Performance Barrier" (KDD 2015).

Here we use ObjectRank on the DBLP dataset as an illustration.

Requirements

• numpy (>= 1.6.2)
• scipy (>= 0.10.1)

Usage

Download the preprocessed data data.tar.gz from here and decompress it:

tar xvf data.tar.gz

Run the following commands to execute the experiments for query answering and learning to rank with model reduction method:

script/answerquery.sh
script/learnrank.sh

After the execution, the experiemental results will be generated in the result folder.

Preprocessing

1. Download the processed DBLP graph file dblp_obj.txt.zip from here, put the unzipped dblp_obj.txt under data/ directory.

2. Compile the code

make -j4

3. Run the following command to generate sampled Personalized PageRank vectors:

bin/ParamPPR data/dblp_obj.txt 2191288 data/sample-params.txt data/sample-vecs/value bin 12

The last parameter 12 is number of threads for computing. You may adjust it based on the number of cores in your machine.

4. Run the following command on a high-memory machine (e.g. 128G) to generate reduce space:

python2.7 src/python/genPriorV.py data/sample-vecs/value bin 1000 3494258 float64 data/basis100.npy

You can also use reduced precision (float32) if the memory is limited (e..g 48G):

python2.7 src/python/genPriorV.py data/sample-vecs/value bin 1000 3494258 float32 data/basis100.npy

5. Run the following command to generate matrix P^{(s)} for each parameter:

bin/GenCSRMatrix data/dblp_obj.txt data/mat/p bin

As discussed in Section 3.3 of the paper, we have:

P(w) = \sum_s=1^7 w_s P^{(s)}`

For the DBLP graph case.