Dual Adversarial Variational Embedding for Robust Recommendation

Qiaomin Yi, Ning Yang, Philip S. Yu

Abstract

Dual Adversarial Variational Embedding (DAVE) is a novel framework designed to enhance the robustness of recommender systems against noisy interaction data. By integrating Variational Auto-encoders (VAEs) with Generative Adversarial Networks (GANs), DAVE learns personalized noise patterns and captures multi-modal user preferences. Experimental results demonstrate that DAVE significantly outperforms existing robust recommendation models across multiple benchmarks, maintaining stability even under high levels of data corruption.

Introduction

In real-world recommendation scenarios, user-item interaction logs are frequently contaminated with noise, such as accidental clicks or interactions driven by temporary trends. Most recommendation models treat these logs as ground truth, resulting in systems that are sensitive to data perturbations. Existing robust methods typically rely on fixed-level noise injection or VAEs with simple Gaussian priors, which fail to capture the personalized nature of noise or the complex, multi-modal characteristics of user preferences.

Methodology

The DAVE framework addresses these limitations through a dual-adversarial structure. It consists of two primary components: UserAVE and ItemAVE, which encode user and item interaction vectors, respectively. By modeling noise as the variance of these inferred distributions, DAVE adaptively learns unique noise levels for each user and item.

The architecture of the Dual Adversarial Variational Embedding (DAVE) framework, showing the integration of UserAVE, ItemAVE, and the Neural Collaborative Filtering network.

To overcome the "single-mode" limitation of standard VAEs, DAVE incorporates adversarial regularization. Auxiliary discriminators ensure that the latent space generated by the inference networks matches a complex prior. This allows the model to represent sophisticated, multi-modal preference distributions—such as a user having distinct interests in disparate genres. The resulting robust embeddings are then processed by a Neural Collaborative Filtering (NCF) network to predict final user-item ratings. The entire system is trained end-to-end, ensuring that recommendation accuracy directly informs embedding learning.

Experimental Evaluation

The model was evaluated on five diverse datasets: Yelp, Digital Music, MovieLens (100K & 1M), and Pinterest. These datasets represent a range of interaction types and sparsity levels. DAVE was compared against several state-of-the-art robust models, including APR, CDAE, and ACAE.

Results and Discussion

DAVE consistently outperformed competing models across various metrics, such as Hit Ratio (HR) and Normalized Discounted Cumulative Gain (NDCG). On the MovieLens 100K dataset, DAVE achieved a notable improvement in NDCG@5 compared to the next best robust method.

Performance comparison of DAVE against state-of-the-art baselines on the MovieLens 100K dataset.

Key insights from the results include:

  • Personalized Robustness: By allowing per-user variance, DAVE acknowledges varying reliability in user histories, treating noise as a learnable feature.
  • Multi-modal Distributions: Adversarial training enables the discovery of distinct interest clusters, which traditional VAEs often merge into a single mode.
  • Noise Tolerance: Stress tests involving intentional data corruption showed that DAVE maintains stable performance even as noise levels increase, widening the gap over its competitors.

Conclusion

DAVE provides a robust framework for recommendation by combining the generative power of VAEs with the flexibility of GANs. While it shows exceptional performance, future work will address challenges in extremely imbalanced datasets and explore the integration of side information, such as user demographics, to further refine personalized noise reduction.

Summary updated

Abstract

Robust recommendation aims at capturing true preference of users from noisy data, for which there are two lines of methods have been proposed. One is based on noise injection, and the other is to adopt the generative model Variational Auto-encoder (VAE). However, the existing works still face two challenges. First, the noise injection based methods often draw the noise from a fixed noise distribution given in advance, while in real world, the noise distributions of different users and items may differ from each other due to personal behaviors and item usage patterns. Second, the VAE based models are not expressive enough to capture the true preference since VAE often yields an embedding space of a single modal, while in real world, user-item interactions usually exhibit multi-modality on user preference distribution. In this paper, we propose a novel model called Dual Adversarial Variational Embedding (DAVE) for robust recommendation, which can provide personalized noise reduction for different users and items, and capture the multi-modality of the embedding space, by combining the advantages of VAE and adversarial training between the introduced auxiliary discriminators and the variational inference networks. The extensive experiments conducted on real datasets verify the effectiveness of DAVE on robust recommendation.
BibTeX
@article{yi2023dual,
  author = {Qiaomin Yi and Ning Yang and Philip S. Yu},
  title = {{Dual Adversarial Variational Embedding for Robust Recommendation}},
  journal = {IEEE Trans. Knowl. Data Eng.},
  year = {2023},
  doi = {10.1109/tkde.2021.3093773},
}
Figure