This repository contains the code related to the research of Ferenc Béres, Róbert Pálovics and András A. Benczúr. If you are interested in our work then you can read the full paper or check out the poster that I presented at the 14th International Workshop on Mining and Learning with Graphs (KDD18).

Please cite our work if you use this code or the Twitter tennis datasets that we collected:

@Article{Béres2018,
author="B{\'e}res, Ferenc
and P{\'a}lovics, R{\'o}bert
and Ol{\'a}h, Anna
and Bencz{\'u}r, Andr{\'a}s A.",
title="Temporal walk based centrality metric for graph streams",
journal="Applied Network Science",
year="2018",
volume="3",
number="32",
pages="26",
}

A former version of our research was presented at the 6th International Conference on Complex Networks and Their Applications. You can find the related codebase on a different branch.

• UNIX environment
• Python 3.5 conda environment with pre-installed jupyter:

conda create -n YOUR_CONDA_PY3_ENV python=3.5 jupyter
source activate YOUR_CONDA_PY3_ENV

• Install the following packages with conda or pip:
  • data processing: pandas, numpy
  • scientific: scipy, networkx, editdistance
  • visualization: matplotlib, seaborn
  • general: sys, os, math, re, json, shutil, operator, collections, multiprocessing, functools, itertools, datetime, pytz

The US Open 2017 (UO17) and Roland-Garros 2017 (RG17) Twitter datasets are available on the website of our research group.

You can also download all related data sets with the following command

bash ./scripts/download_data.sh

After you have downloaded the data sets (see previous section) you can run the experiments with the provided scripts or you can experiment with Jupyter notebooks as well.

You can simply run all the following steps with one script but it will take many time. Look for details in the next section (Calculate centrality scores).

source activate YOUR_CONDA_PY3_ENV
bash ./scripts/run_all.sh

If you read further, you can see the details and instructions related to each task in our experimental setting.

NOTE: In most of the notebooks you have to select from the tennis datasets "rg17" or "uo17".

You can calculate centrality scores for hourly snapshots with the following script.

source activate YOUR_CONDA_PY3_ENV
bash ./scripts/calculate_centrality_scores.sh

It could take several hours to calculate the scores for all parametrizations that is required for the following steps. That is the reason why we use --ExecutePreprocessor.timeout=7200 configuration for the jupyter nbconvert command. If the execution don't finish in 2 hours on your computer then you should increase this time limit.

• Centraity Score Computer: export score files into folder ./data/DATASET_ID/centrality_scores/.

Each implemented centrality measure has a score_id that tries to capture the type and all the parameters of a given method. For example, the score_id is spr_snapshot_12_a0.85_i100 for static PageRank calculated on the last 12 hours of edge history with damping factor 0.85 and 100 iterations. The first part of the score_id always describe the name of the centrality measure:

• tk: temporal Katz centrality (our method)
• ttk: truncated temporal Katz centrality (our method)
• tpr: temporal PageRank
• spr: static PageRank
• hc: static Harmonic Centrality
• nbm: static negative-Beta Measure
• indeg: static indegree
• did: decayed indegree

You can run the experiments with the following script:

source activate YOUR_CONDA_PY3_ENV
bash ./scripts/run_experiments.sh

After you have run the script above, you can find several files related to the experiments (node labels, figures, NDCG@50 scores) in the results/ folder of the repository.

1. Schedule Score Updater: Extract relevant nodes of the Twitter mention networks (daily tennis player accounts). Prepare labels for the supervised evaluation.
2. Predict Tennis Player: Supervised evaluation for daily tennis player prediction based only on network centrality measures. The performance of temporal Katz centrality is compared with several baselines. You must run Centraity Score Computer and Schedule Score Updater notebooks before executing this notebook. You have to specify the list of score_ids (e.g. spr_snapshot_12_a0.85_i100) that you want to evaluate for Normalized Discounted Cumulative Gain (NDCG).
3. Concept Drift: Unsupervised experiment that show that temporal Katz centrality can adapt to changes in the edges distribution.

If you get a "TimeoutError: Cell execution timed out" when you run any of the provided bash scripts then you should increase the execution time limit for the jupyter nbconvert command for the notebook that had broken down. You can do this by using

jupyter nbconvert --to notebook --execute  --ExecutePreprocessor.timeout=HIGHER_TIME_LIMIT BROKEN_DOWN_NOTEBOOK.ipynb

or just execute the given notebook through the Jupyter browser.