Sequential Nature of Recommender Systems Disrupts the Evaluation Process

TL;DR

This paper investigates how the sequential nature of datasets in recommender systems impacts evaluation accuracy, revealing that order effects can influence results by about 1%, especially in competitive scenarios.

Contribution

It introduces adversarial attacks tailored for sequence-aware evaluation, providing a lower bound on information leakage from data order in recommender system assessments.

Findings

Sequence order can affect evaluation outcomes by approximately 1%.

Adversarial attacks can exploit data order to gain additional information.

Evaluation processes are vulnerable to sequence-based manipulations.

Abstract

Datasets are often generated in a sequential manner, where the previous samples and intermediate decisions or interventions affect subsequent samples. This is especially prominent in cases where there are significant human-AI interactions, such as in recommender systems. To characterize the importance of this relationship across samples, we propose to use adversarial attacks on popular evaluation processes. We present sequence-aware boosting attacks and provide a lower bound on the amount of extra information that can be exploited from a confidential test set solely based on the order of the observed data. We use real and synthetic data to test our methods and show that the evaluation process on the MovieLense-100k dataset can be affected by $\sim1\%$ which is important when considering the close competition. Codes are publicly available.