CaDRec: Contextualized and Debiased Recommender Model

TL;DR

A novel hypergraph convolution operator that can select effective neighbors during convolution by introducing both structural context and sequential context is explored and mathematically shows that the imbalance of the gradients to update item embeddings exacerbates the popularity bias, thus adopting regularization and weighting schemes as solutions.

Abstract

Recommender models aimed at mining users' behavioral patterns have raised great attention as one of the essential applications in daily life. Recent work on graph neural networks (GNNs) or debiasing methods has attained remarkable gains. However, they still suffer from (1) over-smoothing node embeddings caused by recursive convolutions with GNNs, and (2) the skewed distribution of interactions due to popularity and user-individual biases. This paper proposes a contextualized and debiased recommender model (CaDRec). To overcome the over-smoothing issue, we explore a novel hypergraph convolution operator that can select effective neighbors during convolution by introducing both structural context and sequential context. To tackle the skewed distribution, we propose two strategies for disentangling interactions: (1) modeling individual biases to learn unbiased item embeddings, and (2) incorporating item popularity with positional encoding. Moreover, we mathematically show that the imbalance of the gradients to update item embeddings exacerbates the popularity bias, thus adopting regularization and weighting schemes as solutions. Extensive experiments on four datasets demonstrate the superiority of the CaDRec against state-of-the-art (SOTA) methods. Our source code and data are released at https://github.com/WangXFng/CaDRec.

Figures (8)

• Figure 1: Illustration of the over-smoothing issue. After two rounds of GCNs, the features of $n_1$, $n_2$, and $n_3$ have propagated throughout the graph, making nodes more indistinguishable.
• Figure 2: Illustration of Individual Bias.
• Figure 3: Illustration of the overall framework, consisting of contextualized and debiased representation learning modules.
• Figure 4: Illustration of item embedding regularization. (a) and (b) denote the well-trained item embeddings without and with regularization in the Yelp2018 dataset, respectively.
• Figure 5: Effect of embedding regularization. The Y-axis represents exponential growth with a base of 2.