OR-AutoRec: An Outlier-Resilient Autoencoder-based Recommendation model
Yuanpeng Hu, Xianmin Wang, Liang Cheng, Jing Li, Di Wu, Yi He · 2022
Deep neural network (DNN) is widely adopted to develop the recommender systems (RSs) in recent years due to its powerful non-linear representation learning ability. So far, various sophisticated DNN-based RSs have been achieved to provide the state-of-the-art recommendation performance. However, most of them ignore the adverse effects caused by outliers (e.g., malicious users). Inevitably, outliers commonly exist in the collected user behavior data. To address this issue, this paper proposes an outlier-resilient autoencoder-based recommendation model, termed OR-AutoRec. Its main idea is to incorporate the Cauchy Loss into an autoencoder to measure the discrepancy between the observed user behavior data and the predicted ones. As such, OR-AutoRec is resilient to outliers owing to the robustness of Cauchy Loss. By conducting extensive experiments on five benchmark datasets, we demonstrate that: 1) our OR-AutoRec is much more robust to outliers than original autoencoder-based model, and 2) our OR-AutoRec achieves significantly better prediction accuracy than both DNN-based and non-DNN-based state-of-the-art models.