TLDR: CE-CDR is a novel framework that improves cross-domain recommendation systems by explicitly modeling causal relationships between user behaviors in different domains. It addresses data sparsity and limitations of existing methods by constructing a causality-aware dataset, learning generalized causal patterns with a unique loss function, and adaptively integrating this causal knowledge into target domain recommendations. Deployed in a real-world platform, it has shown significant improvements in recommendation performance and user engagement.
Recommender systems are everywhere, helping us discover new books, movies, music, and videos. However, these systems often struggle with a common issue: data sparsity. This happens when there isn’t enough interaction data within a single platform or ‘domain’ to accurately understand user preferences. To combat this, Cross-Domain Recommendation (CDR) systems were developed, designed to leverage user data from one domain (the ‘source’) to improve recommendations in another (the ‘target’).
Imagine you’re browsing landscape pictures on a social media platform (the source domain). You see a beautiful park and think, “How wonderful it would be to camp here!” This thought then leads you to watch a camping tutorial video on the same platform (the target domain). This is a classic example of a cross-domain causal relationship, where an action in one area influences behavior in another. By understanding these connections, recommendation quality can be significantly enhanced.
Existing CDR methods often fall into two categories: multi-task learning, which tries to optimize tasks across both domains simultaneously, or feature-based approaches, which use source domain behaviors as additional input. However, these methods face challenges. Multi-task learning can suffer from ‘negative transfer’ if the domains have inconsistent tasks or user behavior patterns. Feature-based methods, while useful, often don’t fully utilize source domain information because they overlook the underlying causal relationships between domains. Simply put, not all source-domain behaviors necessarily influence target-domain preferences.
This is where a new framework, called Causality Enhancement for Cross-Domain Recommendation (CE-CDR), steps in. Proposed by Zhibo Wu, Yunfan Wu, Lin Jiang, Ping Yang, and Yao Hu, CE-CDR takes a novel approach by directly modeling these causal relationships from both a data and model perspective. The core idea is to train a cross-domain representation based on a dataset that explicitly labels causal links from the source to the target domain. This direction has been largely unexplored due to the difficulty of identifying unbiased real causal labels in real-world scenarios.
CE-CDR is structured around three main modules:
Causality Labeling Module (CLM)
The CLM is responsible for creating a high-quality, ‘causality-aware’ dataset. It operates on the psychological assumption that a user’s preference for an item in the source domain will causally influence their preference for a similar item in the target domain. This module uses two types of similarity measures: content similarity (e.g., items belonging to the same category like ‘music’ or ‘sports’) and behavioral similarity (derived from how users interact with items). By combining and calibrating these similarities, CLM identifies strong causal behavior pairs, filtering out noisy data to construct a reliable dataset for training.
Direct Causality Modeling Module (DCMM)
The DCMM is designed to learn how source-domain behaviors truly influence target-domain preferences. It uses a backbone model, like the attention-based SASRec, to predict the probability of a user interacting with a target item due to their source domain behaviors. A key innovation here is the ‘Partial Label Causal Loss’. This special loss function addresses the challenge of imperfect causal labels. It corrects for instances where actual causal pairs might have low similarity (and thus be mislabeled as non-causal) or where the labeling strategy itself introduces bias. By doing so, DCMM can generalize beyond the observed, partially labeled data to uncover unseen causal patterns.
Also Read:
- Beyond Basic Suggestions: A New AI Agent Delivers Explainable and Adaptive Product Recommendations
- Unifying User Interests Across Diverse Online Scenarios with LLM-Enhanced Recommendations
Causality Enhancement Module (CEM)
Finally, the CEM integrates the causal representations learned by DCMM into the target domain recommendation model. It uses a ‘Cross-Domain Self-Attention’ mechanism to allow information exchange between the source and target domain representations, creating more comprehensive user preference profiles. Furthermore, a ‘Cross-Domain Gated Attention’ mechanism adaptively weights the contributions of source and target domain information. This dynamic gating ensures that the model intelligently balances leveraging rich source-domain knowledge with maintaining target-domain relevance, based on specific user features like their behavioral consistency across domains.
The effectiveness of CE-CDR has been demonstrated through extensive experiments on various datasets, including Douban, Amazon, and a large-scale industrial dataset from Rednote (Xiaohongshu). It consistently outperforms existing baseline methods. Notably, when CE-CDR’s causal representation is used as a ‘model-agnostic plugin’ with other methods, it often leads to significant improvements. The framework’s ability to identify unseen causal patterns, even when similarities are low, highlights its generalization capacity.
In a real-world deployment, CE-CDR has been integrated into Rednote, a content-sharing platform serving hundreds of millions of daily active users, since April 2025. Online A/B tests showed statistically significant improvements in key metrics like Click-Through Rate (CTR), diversity, and next-day active users across both video and note recommendation scenarios. The system is also designed for computational efficiency, utilizing a two-phase training scheme, freezing dense parameters during incremental updates, and maintaining a real-time embedding cache to ensure low online latency without sacrificing performance. You can read the full research paper here.
CE-CDR represents a significant step forward in cross-domain recommendation by explicitly modeling causality. Its ability to enhance recommendations by understanding the ‘why’ behind user behaviors, rather than just the ‘what,’ makes it a powerful and practical solution for modern recommender systems.
