Abstract
Cross-Domain Recommendation (CDR) seeks to enable effective knowledge transfer across domains. Existing works rely on either representation alignment or transformation bridges, but they struggle on identifying domain-shared from domain-specific latent factors. Specifically, while CDR describes user representations as a joint distribution over two domains, these methods fail to account for its joint identifiability as they primarily fixate on the marginal distribution within a particular domain. Such a failure may overlook the conditionality between two domains and how it contributes to latent factor disentanglement, leading to negative transfer when domains are weakly correlated. In this study, we explore what should and should not be transferred in cross-domain user representations from a causality perspective. We propose a Hierarchical subspace disentanglement approach to explore the Joint IDentifiability of cross-domain joint distribution, termed HJID, to preserve domain-specific behaviors from domain-shared factors. HJID organizes user representations into layers: generic shallow subspaces and domain-oriented deep subspaces. We first encode the generic pattern in the shallow subspace by minimizing the Maximum Mean Discrepancy of initial layer activation. Then, to dissect how domain-oriented latent factors are encoded in deeper layers activation, we construct a cross-domain causality-based data generation graph, which identifies cross-domain consistent and domain-specific components, adhering to the Minimal Change principle. This allows HJID to maintain stability whilst discovering unique factors for different domains, all within a generative framework of invertible transformations that guarantee the joint identifiability. With experiments on real-world datasets, we show that HJID outperforms SOTA methods on a range of strongly and weakly correlated CDR tasks.
Introduction
Cross-Domain Recommendation (CDR) aims to improve recommendations in a target domain (e.g., Movies) by leveraging data from a source domain (e.g., Books). Traditionally, models attempt to either "align" user features so they appear identical across domains or build a "bridge" to transform features from one to the other.
However, these approaches often treat user representations as inseparable blocks. In reality, some behaviors are consistent across domains (e.g., a general preference for sci-fi), while others are highly domain-specific. Forcing the alignment of these domain-specific factors leads to negative transfer, where noise from one domain degrades performance in the other. Furthermore, existing methods focus on marginal distributions rather than the joint distribution of behaviors, making it mathematically impossible to uniquely recover the true underlying latent factors—a problem known as a lack of identifiability.
Methods: The HJID Framework
The proposed HJID (Hierarchical Subspace Disentanglement) framework addresses these issues through two main pillars:
1. Feature Hierarchy (FH)
Based on the principle that deep neural networks learn general features in shallow layers and abstract, task-specific features in deeper layers, HJID divides the user representation subspace:
- Shallow Subspace Alignment: The model uses Maximum Mean Discrepancy (MMD) to align the activations of initial layers, ensuring that generic, domain-irrelevant patterns are shared.
- Deep Subspace Disentanglement: In deeper layers, the model employs a causal data generation graph to separate latent factors into "stable" (shared) and "variant" (domain-specific) components.
2. Joint Identifiability via Normalizing Flows
To ensure the model can uniquely identify these factors, HJID utilizes Normalizing Flows—a series of invertible (bijective) transformations. By mapping domain-specific factors through these tractable transformations, the model mathematically guarantees that the joint distribution of user representations is identifiable. This allows the framework to capture the correlation between domain-specific behaviors rather than assuming they are independent.
Experimental Evaluation
The model was evaluated using the Amazon Reviews dataset across six pairwise cross-domain tasks, ranging from strongly correlated (e.g., Game-Video) to weakly correlated (e.g., Video-Sport).
Key Results
- Superiority over SOTA: HJID consistently outperformed eight state-of-the-art baselines (including CMF, EMCDR, and CDRIB) across all tasks. The most significant gains were observed in weakly correlated tasks, where traditional alignment methods typically suffer from negative transfer.
- Non-Overlap Success: In experiments where no overlapping users were provided during training, HJID maintained high performance, demonstrating its ability to learn effectively from distributional trends alone.
- Ablation Clarity: Removing either the Feature Hierarchy or the Normalizing Flow components led to a noticeable drop in performance, confirming that both the hierarchical structure and the invertible transformations are critical for successful disentanglement.
Conclusion and Future Work
HJID demonstrates that hierarchical alignment and causal disentanglement are essential for preventing negative transfer in CDR. By ensuring joint identifiability, the model captures complex cross-domain correlations that marginal-focused methods miss.
Future research directions include extending this causal disentanglement to multi-domain contexts (3+ domains simultaneously) and exploring causal discovery to automatically learn relationship structures between domains directly from data, rather than relying on predefined graphs.
Summary updated