Project for Machine Learning on Real World Networks

• The docs folder contains all of the reports - M0, M1, M2, M3
  • The final report is under the M3 folder

• The graph_creation folder contains code that creates the 5-Day Contact Network and analyzes it
  • main.ipynb creates the contact network
  • analysis.ipynb analyzes the contact network
  • The code can be run by executing the code blocks in the .ipynb files
• The link_prediction folder contains code for static link prediction as well as node feature generation for the nodes in the 5-Day Contact Network
  • feature_selection_method1 contains code to generate average locations travelled to per day and average distance travelled per day for each node
  • feature_selection_method2 contains code to extract age, gender, and SAG score for each node
  • feature_combination contains code that combines all features and normalizes them
  • main.ipynb contains the code to train the static link prediction model on the 5-Day Contact Network
  • link_prediction_next_day.ipynb contains the code to perform link prediction on the 6th day using the static link prediction model
  • The code can be run by executing the code blocks in the .ipynb files

• The link_prediction_graph_creation folder contains code that creates the Temporal Contact Networks, performs analysis on those networks, generates node features for nodes in the Temporal Contact Networks, formats the 62 contact networks to be able to be used by the TGN model, and performs an SIR simulation on the Temporal Contact Networks
  • main.py creates the 62 contact networks, one for each day from July 1st, 2020 to August 31st, 2020.
  • feature_selection_method1 contains code to generate average locations travelled to per day and average distance travelled per day for each node
  • feature_selection_method2 contains code to extract age, gender, and SAG score for each node
  • feature_combination contains code that combines all features and normalizes them
  • generate_link_pred_input.ipynb contains code that utilizes the Temporal Contact Networks and their features to generate a data format that can be used by the TGN link prediction framework (convert .gexf files to csv and npy files)
  • simulation.py contains code that uses the Temporal Contact Networks to run an SIR simulation
  • The code can be run by executing the code blocks in the .ipynb files

• The tgn folder contains the code (which is open source and created by the authors or TGN) that is used for the temporal link prediction
  • Original TGN GitHub: https://github.com/twitter-research/tgn
  • Original TGN Paper: https://arxiv.org/abs/2006.10637
  • data contains data used by the framework, which includes edge data and node data
    • Data, such as node features and edge information in csv format, was created in the link_prediction_graph_creation/generate_link_pred_input.ipynb and imported to this tgn/data directory
  • evaluation contains code what evaluates the model. We modified this to plot the ROC AUC curve and generate the confusion matrix that can be used to determine recall
    • We modified this file to generate additional result statistics, such as the confusion matrix, and plot the results
  • modules contains code for the main modules in the TGN framework
    • memory.py - stores memory for the nodes
    • message_function.py - generates messages (information) regarding each node interaction (such as creation of edges)
    • message_aggregator.py - combines a batch of messages to keep one per node
    • memory_updater.py - updates the memory using the current memory and new messages
    • embedding_module.py - generates node embeddings by using the node interaction (such as creation of edges) and the node memory
  • model contains code for the model that is used by the embedding module to create node embeddings
  • train_self_supervised.py contains the code to train and test the temporal link prediction model
    • The model can be trained and tested by running the following command:

      python3 train_self_supervised.py -d data --use_memory --prefix tgn-attn-data --n_runs 10
      

  • generate_visuals.ipynb contains the code to generate charts and graphs such as ROC Curve and Precision-Recall Curve to visualize the results
  • The code can be run by executing the code blocks in the .ipynb files