An Outlier-Resilient Autoencoder for Representing High-Dimensional and Incomplete Data

High-dimensional and incomplete (HDI) data commonly arise in various Big Data-related applications, e.g., recommender systems and bioinformatics. Representation is a learning paradigm to map HDI data into low-dimensional latent space for attracting valuable knowledge and patterns. Currently, deep neural network (DNN) is one of the most popular and successful approaches to represent HDI data due to its powerful nonlinear learning ability. However, previous DNNs-based approaches primarily focused on advancing the sophisticated model structure, neglecting the potential adverse effects of outliers. Unfortunately, outliers usually exist in the collected HDI data. For example, HDI data collected from recommender systems inevitably contain many outlier ratings due to some malicious users. To address this issue, this paper proposes a novel outlier-resilient autoencoder (termed ORA). Its core idea is to design an adaptive Cauchy loss strategy to measure the difference between the observed and predicted data for an autoencoder in representing the HDI data. This strategy leverages a more aggressive Cauchy loss to impose a higher penalty on outlier data with large deviation, while utilizing a smoother Cauchy loss to capture the nuanced, deeper features of HDI data. As such, ORA can dynamically adjust the smoothness of the Cauchy loss during training to handle different levels of data deviation. To evaluate the proposed ORA, extensive experiments are conducted on five benchmark HDI datasets. The results validate that: (1) ORA achieves significantly better representation accuracy than State-of-the-Art DNN- and non-DNN-based models, and (2) ORA possesses higher robustness to outlier data than its peers.