Variational Autoencoder with Stochastic Masks to Solve Exposure Bias in Recommendation (S)

Yingshuai Kou, Neng Gao, Yifei ZhangProceedings/Proceedings of the ... International Conference on Software Engineering and Knowledge Engineering 2023 · DOI

Abstract

In the recommended scenario, an unobserved interaction may exists in two cases: the user is not interested in the item or the user is not aware of the item at all. This phenomenon leads to a serious exposure bias problem in recommendation system. To solve this problem, we propose Masked Variational Autoencoder (MVAE). Firstly, we predict the missing values in the sparse user-item interaction matrix by matrix completion. Then we randomly mask the elements in the obtained matrix and use them as the input to the variational autoencoder. The decoder can reconstruct the user interaction matrix closer to the true distribution and fully exploit the potential preferences of users in uninteracted items. In particular, we use a combined dual VAE to tackle the exposure bias problem from the user side and the item side respectively. Extensive experiments on three real-world datasets also illustrate the effectiveness of MVAE for solving exposure bias in recommendation.

Introduction

In modern recommendation systems, "silence" from a user is inherently ambiguous. If a user hasn't interacted with a specific item, it could be due to a lack of interest, or simply because they were never exposed to it. This ambiguity creates exposure bias, where traditional models often incorrectly treat all unobserved interactions as negative signals. This bias limits the system's ability to discover latent user preferences.

To address this, we introduce the Masked Variational Autoencoder (MVAE). MVAE is designed to "fill in the blanks" of sparse interaction data by combining matrix completion with stochastic masking and a dual-VAE architecture. This approach allows the model to better distinguish between true disinterest and lack of exposure, significantly outperforming existing state-of-the-art models.

Proposed Methodology: The MVAE Framework

The MVAE pipeline consists of three primary stages designed to robustly model user-item interactions:

  1. Matrix Completion (Pseudo-Matrix Generation): The process begins with Singular Value Decomposition (SVD) to estimate missing values in the sparse interaction matrix $R$. This results in a "pseudo-matrix" $R'$ that provides initial estimates for unobserved interactions based on global patterns.
  2. Stochastic Masking: To prevent the model from overfitting to the SVD's initial guesses, random elements of the pseudo-matrix are masked (set to zero). This stochastic masking acts as a regularizer, forcing the model to learn the underlying distribution rather than just memorizing the input.
  3. Dual VAE Architecture: The core of MVAE utilizes two Variational Autoencoders (VAEs) in parallel:
  4. uVAE (User-based): Captures user-specific preference patterns by processing the matrix row-wise.
  5. iVAE (Item-based): Captures item-specific exposure patterns by processing the matrix column-wise. The final prediction $\hat{R}$ is a weighted combination of the outputs from both decoders, ensuring a balanced view of both user taste and item characteristics.
The MVAE architecture: from interaction matrix to final reconstruction via SVD, masking, and dual VAEs.

Experimental Evaluation

We evaluated MVAE on three diverse real-world datasets: Movielens1M (96.89% sparsity), Yelp (99.73% sparsity), and Pinterest (99.73% sparsity). These datasets represent varying scales and domains, providing a rigorous test for the model's ability to handle extreme sparsity.

Key Results

MVAE was compared against several baseline models, including Bayesian Personalized Ranking (BPR), Neural Collaborative Filtering (NCF), and Mult-VAE.

Performance comparison of MVAE against baseline models across three datasets.

As shown in the results table, MVAE consistently achieves superior performance across almost all metrics, including F1-score and Normalized Discounted Cumulative Gain (NDCG). Notably, on the Pinterest dataset, MVAE improved the NDCG@10 by approximately 9.375% over the best-performing baseline. The model's advantage becomes even more pronounced as the number of recommended items ($K$) increases, demonstrating its effectiveness at surfacing relevant "unseen" items.

Discussion and Insights

The success of MVAE stems from two key design choices:

  • SVD as a Guide: By using SVD as a "hint" rather than a final output, MVAE leverages the generative power of VAEs to refine these initial guesses into robust probability distributions.
  • The Power of Masking: Randomly hiding data during training forces the model to reconstruct the interaction matrix closer to its true underlying distribution, effectively mitigating the impact of observed (biased) data.
  • Dual Modeling: Ablation studies confirmed that the dual approach (uVAE + iVAE) outperforms single-stream models, suggesting that user preferences and item exposure are complementary signals.

Conclusion and Future Work

MVAE effectively addresses exposure bias by combining matrix completion with a generative dual-VAE framework. While the model shows slight sub-optimality for $K=1$ on extremely large item sets like Yelp, its overall performance is state-of-the-art. Future research will explore incorporating temporal dynamics to account for evolving user interests and item popularity over time.

Summary updated

Abstract

In the recommended scenario, an unobserved interaction may exists in two cases: the user is not interested in the item or the user is not aware of the item at all. This phenomenon leads to a serious exposure bias problem in recommendation system. To solve this problem, we propose Masked Variational Autoencoder (MVAE). Firstly, we predict the missing values in the sparse user-item interaction matrix by matrix completion. Then we randomly mask the elements in the obtained matrix and use them as the input to the variational autoencoder. The decoder can reconstruct the user interaction matrix closer to the true distribution and fully exploit the potential preferences of users in uninteracted items. In particular, we use a combined dual VAE to tackle the exposure bias problem from the user side and the item side respectively. Extensive experiments on three real-world datasets also illustrate the effectiveness of MVAE for solving exposure bias in recommendation.
BibTeX
@inproceedings{kou2023variational,
  author = {Yingshuai Kou and Neng Gao and Yifei Zhang},
  title = {{Variational Autoencoder with Stochastic Masks to Solve Exposure Bias in Recommendation (S)}},
  booktitle = {Proceedings/Proceedings of the ... International Conference on Software Engineering and Knowledge Engineering},
  year = {2023},
  doi = {10.18293/seke2023-047},
}
Figure