Synthetic disjoint and overlapping dynamic attributed networks with ground-truth information
Description
1. Synthetic network 1: Using DANCer, we created 12 dynamic undirected attributed datasets (10 timestamps) with homogeneous communities. Over time, nodes increase from 256 to 1868, communities from 5 to 10, and edges from 1000 to 11905. 2. Synthetic network 2: 200-node graphs over 20 snapshots, starting as two 100-node groups. At a random step, 40% (Dataset 1) or 80% (Dataset 2) of nodes migrate to a new community. Generated via SBM (0.3 intra-community, 0.1 inter-community probability). 3. Synthetic network 3: Ten timestamps with growing/shrinking communities (three groups of 80, 90 and, 100 nodes), via a degree-corrected SBM for the network structure. The proposed attributes assess strongly assortative (Dataset 1), weakly assortative (Dataset 2), and disassortative (Dataset 3) structures. 4. Overlapping synthetic networks: Using Greene's benchmark, we create 11 overlapping datasets with 10 snapshots. Each network has 250 nodes, overlapping nodes from 0 to 50, and overlapping memberships from 2 to 25. Specific kinds of changes in communities of a dynamic network were simulated: - Birth and death: Two birth events per time step and two death events per time step. Dataset 1 has no overlapping nodes; Dataset 2 has five overlapping nodes and two overlapping memberships; Dataset 3 has ten overlapping nodes and two overlapping memberships; Dataset 4 has 25 overlapping nodes and two overlapping memberships; Dataset 5 has 50 overlapping nodes and two overlapping memberships; Dataset 6 has five overlapping nodes and five overlapping memberships; Dataset 7 has five overlapping nodes and ten overlapping memberships; Dataset 8 has five overlapping nodes and 25 overlapping memberships. - Merge and split: Two merge events per time step and two split events per time step. Dataset 9 has five overlapping nodes and two overlapping memberships. - Expansion and contraction: Two expansion events per time step and two contraction events per time step, with a rate of 0.1. Dataset 10 has five overlapping nodes and two overlapping memberships. - Node switching between communities: A probability of 0.2 for a node to switch its community membership. Dataset 11 has five overlapping nodes and two overlapping memberships. 5. Real-world network: The crime network comprises individuals who committed crimes in Chile between 2019 and 2021. The edges of the graph denote joint participation in a case. To construct the timesteps, the date of the crime was used to establish three-month intervals, where each period is non-overlapping. The attributes considered in this dataset are the crimes committed by each subject, where each column designates a specific crime and a value of 1 indicates that the individual committed the corresponding crime. Robbery in an uninhabited place, robbery with violence, and homicide in a fight, fall under the scope of this network, which includes a total of 26 crime types. In this dataset, the attribute values are time-varying.
Files
Steps to reproduce
The datasets were used in the paper "Detecting disjoint and overlapping communities in temporal node-attributed networks". Synthetic network 1 were built with the synthetic benchmark proposed by Largeron et al. (2017), doi:10.1007/s10115-017-1028-2, with the code available at https://perso.univ-st-etienne.fr/largeron/DANC_Generator/. Synthetic network 2 was based on Sheikholeslami and Giannakis (2018), doi:10.1109/TSP.2018.2871383. Synthetic network 3 was based on Tang et al. (2020), doi:10.1007/s00180-019-00909-8, with the addition of time steps. Overlapping synthetic networks were based on Greene et al. (2010), doi: 10.1109/ASONAM.2010.17, with the code available at http://mlg.ucd.ie/dynamic/. About attributes: Attributes for each dataset were added as follows. For a network of $c$ communities, each one was assumed to have a strong correlation with $h$ binary attributes and a weak correlation with $h \cdot (c-1)$ binary attributes. We use $h = 1$ and $h = 3$ for our tests. The probabilities of having a strong correlation (pin) were varied from 0.5 to 1, while the probabilities of having a weak correlation were set with 0.05 or 0.1 (pout). For example, a folder named pin06pout005 indicates that the probability of having a strong correlation is 0.6 and the probability of having a weak correlation is 0.05. About files: Each dataset contains ten folders, each with a different seed. Inside each folder: - .gml files represent the topology for each time step. - The files starting with the letters GT* and OGT* are the ground truths for the disjoint and benchmarks, respectively. - The file starting with K* is the number of communities for each time step for the disjoint benchmarks. - The file finalizing in *Index.txt contains the node numbers for each time step needed to track the creation and deletion of nodes, and it corresponds with attribute values. - The attributes are in .csv format. - For overlapping synthetic networks, files with extension .comm, .edges, .stats, and .timeline, are format files generated by the benchmark. For example, Synthetic network 1/Synthetic network 1 Dataset 1/Data122NS3/ refers to the seed 3 of the first dataset of synthetic network 1, where: - Data122NS3Index.txt are the indices of nodes for all timestamps of seed 3. - Data122NS3t4.gml contains the topology for time step 4 of seed 3. - GTData122NS3t4 contains the ground truth for time step 4 of seed 3. - KData122NS3.txt contains the number of communities for all timestamps of seed 3. The maximum of communities in this case is 9, for timestamps 9 and 10. - pin06pout005/Data122NS3H1Attrt4.csv contains the attributes for time step 4 of seed 3, where nodes inside a community have one binary attribute (h=1) with a probability of 0.6 to have the same value as other nodes of the same community. The nodes inside a community have another eight binary attributes (1 * (9-1)) with a probability of 0.05 to have the same value as other nodes of other communities.