A Debiasing Autoencoder for Recommender System

The deep neural network (DNN)-based recommender system (RS) has drawn much attention recently and provided state-of-the-art results. Although many DNN-based RSs have been achieved, most focus on inventing a sophisticated DNN with a single-metric loss function to fit user behavior data. However, user behavior data are commonly collected from numerous users in complex scenarios, making various biases and outliers (outliers can be seen as the special bias) widely exist in the data. Unfortunately, prior DNN-based RSs only considered rather fragmented biases and lacked a comprehensive solution. To fill this gap, this paper proposes an AutoRec++ model to comprehensively address the various biases existed in user behavior data. Its main idea is to employ different combinations of preprocessing bias (PB) and training bias (TB) as well as
$L_{1}$
-norm and
$L_{2}$
-norm to form a multi-metric loss function-oriented Autoencoder. As such, AutoRec++ possesses the multi-merits of the PB’s and TB’s debiasing ability, the
$L_{1}$
-norm’s robustness, and the
$L_{2}$
-norm’s stability. By conducting extensive experiments on five benchmark datasets, we demonstrate that: 1) the incorporation of PB and TB can significantly boost Autoencoder’s prediction accuracy and computational efficiency without structural change, and 2) our AutoRec++ achieves better prediction accuracy and robustness than both DNN-based and non-DNN-based state-of-the-art models. Besides, our AutoRec++ is more effective in processing sparser user behavior data. Our code is available at the link:
https://github.com/wudi1989/AutoRec_Plus
.