Paulpanther / info-flow

Robust Source Detection

Part of the Information Flow Seminar in the Winter Term 2021/2022 at HPI.

In this repository, you can find:

1. An open-source python implementation of Rumor Centrality in rumor_centrality.rumor_detection.py
2. Easy graph generation using networkx in rumor_centrality.graph_generator.py
3. Easy infection simulation using ndlib in rumor_centrality.graph_simulations.py
4. Notebooks and code to run experiments with different combination of graphs, spreading behavior, and source detection algorithms (mainly TopologyMatrix.ipynb)
5. Notebooks and code to experiment with missing values in source detection (missing_experiment.py and EDA of Missing Data Experiments.ipynb)
6. Notebooks and code to experiment with multiple sources (multiple_centers_experiment.py and the Multiple Rumor Centers notebooks)

Setup

To run the code here, you need at least Python 3.6 and then install the requirements in requirements.txt. Usage of a virtual environment is encouraged.

pip install -r requirements.txt

Comparison of Centrality Metrics in Different Scenarios

The Notebook TopologyMatrix.ipynb visualizes a comparison between Rumor Centrality, Distance Centrality, Betweenness Centrality, and Jordan Centrality on Small World, Scale Free, Synthetic Internet and US Power Grid graphs infected with SI, SIS and SIR to the infection sizes of 10, 100, 200, 300, 400, 500, 600, 700, 800 and 900 nodes. The predictions are evaluated using Mean Diameter Normalize Hop Distance.

The second cell of the Notebook contains the parameters for the Dynamics, Predictiors, Infections Counts, Graphs and Amount of Repetitions. All parameters can be changed there and new dynamics, graphs and predictors can be added.

The following cells describe the process of generating, simulating and predicting. Because the process take a long time, it can be manually subdivided into smaller steps and the partial results can be stored and loaded using the save and load functions.

The last cell renders the matrix. Here some visualization parameters can be changed, like which data to show, the tick size and data range.

Missing Node Experiments

While collecting the status of nodes of an information spread, false negatives can occur. This leads to actually infected people, actually congested traffic junctions or similar not being represented as a node in the infection graph (there obviously can be false positives as well, but we did not experiment with that).

The experiments of testing the impact of missing nodes for the informaction source detection follow the following process:

1. Simulate an infection using the SI Model in the chosen network until 30% of all nodes are infected
2. Use all four metrics (Rumor Centrality, Jordan Centrality, Distance Centrality, Betweenness Centrality) to predict the source of the simulated infection
3. Remove a certain percent of infected nodes, reconnect the neighbors of the removed node to a clique and repeat the source prediction step on the resulting reduced and reconnected infection graph
4. Measure the distance of all predicted sources to the original source of the simulation

Each step of the process creates some data usefull for our research:

1. Creates a baseline infection graph
2. Creates a baseline prediction (i.e. 0% of infected nodes removed)
3. Creates a reduced infection graph and source predictions for that graph
4. Creates distances form predicted sources to original source, thus making all source predictions compareable

The described process and the data created can be reproduced with the missing_experiment.py script:

python3 missing_experiment.py graph output_dir diffusion_dynamic 

The options for graph are: synthetic_internet_100, synthetic_internet_1000, scale_free_100, scale_free_1000 , us_power_grid, and internet (Note that we only use synthetic_internet_1000, scale_free_1000 and us_power_grid).

You can freely choose the output directory, if it does not exist, the script will create it.

The options for diffusion_dynamic are: si, sis, and sir (Note that we only use si).

The script will create a folder for the dynamic you chose and a sub-folder for the chosen graph ( eg. output_dir/si/scale_free_100). In this folder you will find the results of the computation. To better separate the files of different experiments, the results are named after the following schema: {data_name}__config_graph_{graph}_nodes_{graph_size}_samples_{sample_size}_ There are 5 files per run, all identified by the {data_name} attribute:

1. bar__[...]: A hist plot for each of the four metrics. Each hist visualizes the frequencies of hop distances from predicted source to original source, dependent on how many nodes were removed. Hop Distance on x, missing percent as data.
2. stacked_bar__[...]: Contains the same data as bar__[...] but visualizes as stacked bar plot, with missing percent on x and the hop distance distribution as plotted data.
3. main_ref_graph__[...]: This is a pickle file containing the networkx graph object of the base graph, the infection simulations were run on. This is important, as synthetic graphs are not fixed to a seed, and change for each run of the script.
4. results__[...]: The main experiment data. A pickle file, containing a dictionary, mapping percent missing to a list of all sample results. Each sample result contains the original sources (real_centers), the reduced infected graph (ex_graph), the nodes that were removed to create ex_graph (removed_nodes) and finally predicted_centers, which contains a dictionary with a key for each metric. In this dictionary each metric maps to a list of all the nodes, that were detected as potential information flow sources.

You can use the EDA of Missing Data Experiments.ipynb notebook to read in the results of the script, without parsing manually. This notebook also contains logic to normalize and visualize our results, exactly the same way we used for creating our report.

Multiple Rumor Centers

An infection can have more than one source. We want to try to infer those. For that, we first partition the graph in k subgraphs, and then apply the metrics on those. Specifically, we follow those steps:

1. Simulate an infection using the SI Model in the chosen network until the desired number of nodes are infected on from k many sources (making sure the resulting graph is connected)
2. Partition the graph into k
3. Use all four metrics (Rumor Centrality, Jordan Centrality, Distance Centrality, Betweenness Centrality) to predict the source of the simulated infection in each partition
4. Measure the distances from the predictions to the original sources

The approach is outlined in Multiple Rumor Centers - Overview.ipynb. Most of the code for this can be found in rumor_centrality.graph_clustering.py

The first 3 steps can be carried out by using the multiple_centers_experiment.py script:

python multiple_centers_experiment.py <cluster_numbers> <output_dir>

where cluster_numbers is a comma seperated list of clusters, for which the experiment should be run. Example usage:

python multiple_centers_experiment.py 2,3,5,7 results

This runs the experiment on the graphs synthetic_internet_10000, scale_free_10000, and us_power_grid repeating each experiment 100 times.

The output will be written to the <output_dir> folder in form of a pickle file containing a Dictionary for each run containing:

• the hops (list of hops)
• graph_normalized_hops
• diameter
• num_infection_centers
• max_infected_nodes
• predictions (i.e. the predicted source nodes, also a list)
• ground_truths (i.e. the original sources, also a list)
• the actual networkx graph
• as well as metric

Instead of the script, you can also use the notebook Multiple Rumor Centers - Experiment.ipynb to carry out the experiment. The main method for this can also be found in rumor_centrality.experiment.py

Using the pickled results file, you can use the Multiple Rumor Centers Analysis.ipynb notebook to analyze your results. This performs the necessary data cleaning and preparation as well as visualization to get the same results as seen in our report.