Studying Graph Based Recommendation Systems under Adversarial Pressure

CS224W @ Stanford, Fall 2018

---

Overview

IMPORTANT: because SNAP.py uses Python 2.7, we use Python 2.7 in this project as well.

This repo houses code that experiments with how graph-based recommender systems react under Adversarial pressure. We use a variant of the Pixie algorithm to recommend items to users on bipartite networks. Throughout our code we use the SNAP library for common graph operations.

Directory Structure

• gbra: contains all our core code. Any miscellaneous scripts and documentation files live in the outer level of the repo.

  • gbra/recommenders: contains the recommender system black box. Recommender systems in this module can be instantiated on a graph, trained, evaluated, and asked for predictions.
  • gbra/attackers: contains different classes of attackers.
  • gbra/data: contains routines for fetching different datasets and instantiating them as graphs.
  • gbra/feature_extraction: contains classes and routines for extracting features from graphs.
  • gbra/util: contains utility functions such as EIGraph, an abstraction for Entity-Item graphs, and some useful asserts.

• exp_*.py: experiment files. To maintain some order on the repo, make a new python file for each separate experiment you plan to run, then run it as python exp_<experiment name>.py from gbra/. Add all graphing code to its corresponding experiment file.

• Utility logic for each submodule lives in gbra/util/<submodule>_utils.py.

Quick Start

1. Create a Python 2 virtual environment and activate it.
2. Install SNAP from within that virtual environment
3. python exp_sample.py
4. Profit