Recommender systems are widely used to help users discover content of interest. However, due to their reliance on observational user–item interaction data, they often suffer from data bias. Such biases primarily stem from non-random exposure and users’ self-selection behavior, which distort the data distribution and lead to suboptimal performance of recommendation models. Existing debiasing methods, especially those based on loss reweighting strategies, have shown promising empirical results but still lack solid theoretical guarantees. In particular, they struggle to handle the complex, diverse, and often unidentifiable types of bias encountered in real-world scenarios. In this article, we revisit the problem of unbiased recommendation from the perspective of data bias and propose a unified debiasing framework that mitigates the effect of bias by aligning the distribution of training data with that of unbiased data collected under randomized exposure. We provide a thorough analysis of the theoretical limitations of existing reweighting methods, and we further propose a principled method, mEta Representation Alignment (ERA) , aiming to alleviate the inconsistency between user and item features under different distributions. Extensive experiments on real-world and semi-synthetic datasets demonstrate the effectiveness of ERA.

User interactions are often driven by latent, unobservable intentions, which are crucial for understanding and predicting behavior in recommendation systems. Previous work has attempted to learn these latent intentions through auxiliary information or clustering, yet uncertainty of interactions and highly entangled intentions still impact recommendation effectiveness. Thus, probabilistically modeling and disentangling users’ latent intents can better capture these high-dimensional behavior patterns. We propose Contrastive Intent-Disentangled Variational AutoEncoder (CIDVAE), a novel method to model and disentangle complex latent intentions in Sequential Recommendation (SR). CIDVAE addresses major SR challenges including evolving user intention uncertainty, intention disentanglement, and data sparsity. It comprises two components: an intent-disentangled variational autoencoder that separates distinct intentions from item sequences, and a contrastive learning module that refines disentanglement by maximizing mutual information between latent representations of the same intention. Experiments on four real-world datasets demonstrate CIDVAE’s superior performance and robustness in SR tasks.

Data sparsity poses a significant challenge for recommendation systems, prompting the research of Cross-Domain Recommendation ( CDR ). CDR aims to leverage more user-item interaction information from source domains to improve the recommendation performance in the target domain. However, a major challenge in CDR is the identification of transferable features. Traditional CDR methods struggle to distinguish between the various features of users, including domain-invariant features that are effective for feature transfer and domain-specific features that are detrimental to cross-domain information transfer. In this article, we aim to disentangle domain-invariant features and domain-specific features and effectively utilize these different features. This enables effective domain-to-domain information transfer by only transferring domain-invariant features while still considering the role of domain-specific features within their respective domains. Based on the superiority of graph structural feature learning and disentangled represent learning, we propose \(\mathbf{DMGCDR}\) —a model that learns D isentangled user feature representations and constructs a M ulti- G raph network for bidirectional knowledge transfer of shared features for CDR . Specifically, we designed two regularization terms to disentangle domain-invariant features and domain-specific features. Subsequently, we established a multi-graph convolutional network to enhance domain-specific features within single-domain graphs and transfer domain-invariant features across cross-domain graphs. Our approach also includes designing feature constraints to enhance the combination of features derived from different graphs and to uncover potential correlations among them. Extensive experiments on real-world datasets have demonstrated that our model significantly outperforms state-of-the-art CDR approaches.

The recommendation system is subject to various biases, resulting in different training and testing data distribution. Most previous work either relies on a part of uniform data to guide model training which is difficult to obtain, or trains without any use of uniform data. However, not using uniform data may result in the inability to observe user’s real behaviors, leading to the presence of confounding factors, which harms the performance of unbiased recommendations. In this work, we proposed a novel method IML to perform debiasing recommendations by leveraging only the statistical characteristics of uniform dataset and training data. We use Invariant Meta-Learning(IML) to learn invariant features that remain insensitive to distributional changes. Finally, we propose a sample hard-aware weighted method to enhance training. Extensive experiments on real-world datasets demonstrate the effectiveness of IML.

In recommender systems, learning high-quality user and item representations is crucial for predicting user preferences. However, there are various confounding factors in observational data, resulting in data bias, which hinders the learning of user and item representations. Recent work proposed to use uniform data to alleviate bias problem. However, these methods fail to learn pure representations for unbiased prediction, which are not affected by confounding factors. This paper introduces a novel disentangled framework, named CDLRec, for learning unbiased representations, leveraging uniform data as supervisory signal for disentangling. Furthermore, to address the scarcity problem of uniform data, the contrastive learning is utilized to implement disentanglement by providing augmented samples. Specifically, two contrastive strategies are designed based on different sampling ways for positives and negatives. Extensive experiments are conducted over two real-world datasets and the results demonstrate the superior performance of our proposed method.