• The paper addresses the issue of item recommendation, a task typically solved using personalized ranking objectives like Bayesian Personalized Ranking (BPR).
• It highlights the vulnerability of models like Matrix Factorization (MF) optimized with BPR to adversarial perturbations, affecting their robustness and generalization.

Motivation

• Adversarial Vulnerability: Advances in adversarial machine learning reveal that many classifiers, including those in information retrieval (IR), are fragile and can be easily misled by adversarial examples.
• Adversarial Examples in IR: No prior study has explored adversarial learning for information retrieval. This paper aims to fill that gap by applying adversarial learning to item recommendation.

APR Overview :

• Framework: APR enhances BPR by introducing adversarial training. It aims to make the model robust by defending against adversarial perturbations.
• Minimax Game: The training process involves minimizing the BPR objective while countering an adversary that adds perturbations to maximize the same objective.
• Implementation: The authors implement APR on MF by adding adversarial perturbations to user and item embedding vectors.

Matrix Factorization

• Definition: MF is a fundamental and effective model in recommendation systems, representing users and items as embedding vectors. The core idea of MF is to estimate a user preference on an item as the inner product between their embedding vectors

• Predictive Model: The preference of a user u for an item i is estimated as the inner product of their embedding vectors.

  

  

Bayesian Personalized Ranking

• Objective: BPR is a pairwise learning-to-rank method designed to optimize recommender models by maximizing the margin between observed and unobserved interactions.
• Assumption: Observed interactions should be ranked higher than unobserved ones.
• Optimization: Typically done using Stochastic Gradient Descent (SGD).

MF-BPR is vulnerable to Adversial Noise