Academic Network Project

ANP Core

The anp_core module is the foundation of the ANP system: it manages datasets, builds and extends the academic (hindsight) infosphere, and offers a range of utilities for data processing and analysis within the Academic Network Project. Key functions include loading datasets, assembling the infosphere, expanding its scope, and other essential processing tools.

A standard execution flow is:

1. Import the AMiner dataset into PyG and parse its contents.
2. Use anp_infosphere_creation to construct the infosphere, with options to split the work into multiple parts for parallel processing.
3. For each part, call anp_infosphere_expansion_caller to run the anp_expansion routines.
4. Combine all expanded segments by running anp_infosphere_builder.

---

ANP NN

The anp_nn package contains Graph Neural Network models designed for prediction within the ANP framework. It currently includes:

• Co-Author Prediction: A model based on Heterogeneous Graph Transformers (HGT) to predict future collaborative links among researchers.

• Synthetic Ground-Truth Generation: Creates simulated interaction datasets using predefined recommender strategies. It:

  1. Trains a Recommender-Neutral User (RNU) model on historical network data.
  2. Simulates various recommendation approaches (e.g., no infosphere, hindsight infosphere, top-paper, top-paper × topic, LightGCN).
  3. Generates labeled interaction pairs according to each approach’s logic.

This addition allows for controlled benchmarking of models and recommender detection methods against known synthetic ground-truths.

For the experiments in the paper, a learning rate of 0.00001 and 50 sampled edges for minibatch creation were used as hyperparameters. The other hyperparameters remain fixed and can be found in the code.

Experimental Procedure

To run the experiments, after parsing the dataset and generating the hindsight infosphere, follow the steps below.

1. Train the RNU using the script anp_link_prediction_co_author_hgt.py and generate the corresponding weights.
2. Create a Ground Truth (GT) for each available infosphere by running create_GT/create_gt_opt.py.
3. Run the experiments for each generated GT and each infosphere using the script anp_link_prediction_co_author_hgt_GT.py, applying the parameters specified here and in the paper.

---