🤖 AI

CoRCi: Cross-Reconstruction of Coherent Interests Modeling in Cross-Domain Sequential Recommendation

CoRCi is a novel cross-domain sequential recommendation framework that addresses data sparsity and interest coherence issues by employing a cross-reconstruction mechanism to generate mixed-domain representations and a FocalNCE loss function to strengthen domain-invariant alignment, thereby outperforming state-of-the-art methods.

Original authors: Qingtian Bian, Tieying Li, Marcus de Carvalho, Jiaxing Xu, Hui Fang, Yiping Ke

Published 2026-08-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Qingtian Bian, Tieying Li, Marcus de Carvalho, Jiaxing Xu, Hui Fang, Yiping Ke

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to guess what your friend wants to eat next. If you only know they love pizza, you might suggest more pizza. But what if you also know they just bought a fancy coffee maker and some gourmet beans? Suddenly, you realize they aren't just a "pizza person"; they are a "foodie" with a specific, evolving taste for high-quality ingredients, whether that's Italian or Italian-inspired coffee. This is the heart of Recommender Systems: the digital matchmakers that try to predict what you'll like next based on what you've done before.

Usually, these systems work in silos. A movie app only knows your movie history, and a grocery app only knows your food history. But people are messy; we jump between categories. We buy a grill pan, then a bag of coffee, then a snack. Our interests flow across these boundaries. The challenge scientists face is Cross-Domain Sequential Recommendation. It's the art of connecting the dots between your "kitchen" life and your "food" life to understand your true, evolving taste. The big question is: how do we teach a computer to see that a grill pan and a bag of coffee are part of the same story, without getting confused by the different types of items?

This is where a new paper called CoRCi (which stands for Cross-Reconstruction of Coherent Interests) steps in. The researchers, a team from universities in Singapore and China, noticed that previous attempts to connect these different "worlds" were a bit clumsy. They were like trying to stitch two different fabrics together by just gluing the edges; the pattern inside the fabric would get distorted.

The authors argue that the old way of doing things treats your "kitchen" history and your "food" history as two separate stories that just happen to be told at the same time. They found that when you force a computer to learn from these mixed stories using the old methods, it gets confused. It starts seeing the differences between the domains (like "this is a pan" vs. "this is coffee") as more important than the similarities (like "I am cooking a healthy meal"). This causes the computer to lose the "coherence" of your interests—the smooth, logical flow of your personality across different topics.

To fix this, CoRCi introduces a clever two-step dance. First, instead of building a new brain from scratch to understand the mixed-up story, it takes the "brain" it already built for the specific stories (the kitchen-only brain and the food-only brain) and uses a special bridge called Cross-Reconstruction. Imagine you have two translators who speak different languages perfectly. Instead of forcing them to learn a third, mixed language from scratch, you let them talk to each other to create a shared understanding of the story they are both telling. CoRCi uses this "cross-attention" to blend the specific details into a unified, coherent picture of your interests.

Second, the paper introduces a new scoring rule called FocalNCE. Think of this as a strict teacher who knows that students often make mistakes with things that look very similar. In the world of recommendations, items from the same category (like two different types of coffee) are very similar, so a computer might prioritize one because it looks like the other. FocalNCE punishes the computer harder when it confuses items from the same category, forcing it to pay closer attention to the unique, cross-domain connections that actually matter. It's like telling the computer, "Don't just pick the coffee because it looks like the other coffee; pick it because it fits the whole story of your cooking adventure."

The researchers tested this idea on four real-world datasets, including massive collections of reviews from Amazon (covering things like food, kitchenware, beauty products, and electronics) and Douban (movies, books, and music). They found that CoRCi consistently beat the best existing methods. In fact, on the Amazon Food-Kitchen dataset, it improved the ability to predict the next item by about 5% to 11% compared to the previous champions. On the Beauty-Electronics dataset, the gains were even more impressive, reaching up to 11%.

The paper also ran some "what-if" experiments to prove why their method works. They tried removing the special bridge (Cross-Reconstruction) or the strict teacher (FocalNCE), and the performance dropped. They even tested it on a dataset with three domains (movies, books, and music) where users didn't necessarily interact with all three, and CoRCi still won.

In short, the paper suggests that to truly understand what a user wants next, we shouldn't just mash their different histories together and hope for the best. Instead, we should carefully reconstruct their interests to keep the story coherent, and use smarter scoring rules to ensure the computer focuses on the unique connections between different parts of their life. It's a small but significant step toward making our digital assistants feel less like rigid databases and more like friends who really "get" us.

Technical Summary: CoRCi

Problem Statement
Cross-Domain Sequential Recommendation (CDSR) aims to alleviate data sparsity by transferring dynamic user interests across related domains. A central premise is that certain user interests are domain-invariant and evolve coherently across multiple domains. However, existing CDSR methods face two primary challenges:

Loss of Interest Coherence: Current approaches often merge domain-specific sequences into a mixed-domain sequence but employ separate encoders trained with per-domain loss aggregation (Dual-Target strategy). This workflow fragments user interests, failing to preserve the inherent coherence of domain-invariant patterns, particularly when query-target pairs in Seq2Seq models originate from different domains.
Domain Discrepancies: Most models learn mixed-domain representations from scratch using separate encoders. This introduces significant domain gaps, as queries and ground-truth targets often belong to different domains, degrading performance. Furthermore, standard contrastive losses in mixed-domain settings may introduce intra-domain bias, where negative samples from the same domain as the query are too similar, skewing the representation space.

Methodology: CoRCi
The authors propose CoRCi (Cross-Reconstruction for Coherent Interests), a dual-target CDSR framework designed to bridge these domains while preserving interest coherence. The architecture consists of three main components:

Specific-Domain Self-Attention Encoders (SAE): Instead of training a single encoder from scratch on mixed data, CoRCi first encodes specific-domain sequences ( $T_A$ and $T_B$ ) independently using dedicated Self-Attention Encoders. This preserves the initial domain-specific patterns.
Cross-Reconstruction (CR) Module: To reconnect dispersed domain-invariant interests, CoRCi employs a Cross-Reconstruction approach. It generates mixed-domain representations ( $H_M$ $H_{M}$ ) directly from the pre-encoded specific-domain representations ( $H_A, H_B$ $H_{A}, H_{B}$ ) using a cross-attention mechanism.
- The query ( $Q$ ) is the mixed-domain embedding ( $E_M$ ), introducing mixed-domain positional information.
- The key ( $K$ ) and value ( $V$ ) are the sum of the pre-encoded specific-domain representations ( $H_A + H_B$ ).
- This allows the model to reconstruct a coherent mixed-domain view without learning it from scratch.
Cross-Attention Transferors (CAT): The reconstructed mixed-domain knowledge is transferred back to the specific-domain representations via Cross-Attention Transferors. A stop-gradient operation is applied to the mixed-domain representation ( $\hat{H}_M$ ) during this transfer. This ensures that the mixed-domain modeling acts as a task-agnostic backbone, preventing specific-domain supervision from distorting the coherence of domain-invariant interests.
FocalNCE Loss: To mitigate intra-domain bias in the mixed-domain objective, the authors introduce FocalNCE. This integrates Focal Loss into the standard InfoNCE objective.
- In the mixed-domain setting, negative samples are drawn from both specific domains.
- FocalNCE assigns higher penalties to negative samples drawn from the same domain as the query (positive item).
- This forces the model to focus on harder, cross-domain negatives, thereby strengthening domain-invariant alignment and suppressing domain-specific noise.

The final objective function is a linear combination of the specific-domain losses ( $L_A, L_B$ ) and the mixed-domain loss ( $L_M$ ), weighted by a hyperparameter $\beta$ .

Key Contributions

Cross-Reconstruction Approach: A novel method that generates mixed-domain representations from pre-encoded specific-domain sequences via cross-attention, effectively reconnecting coherent domain-invariant interests that are otherwise dispersed.
FocalNCE: A new loss function embedding Focal Loss into the mixed-domain InfoNCE objective to mitigate intra-domain bias and prevent the fragmentation of coherent interests.
Comprehensive Evaluation: Extensive experiments on four real-world datasets (Amazon Food-Kitchen, Beauty-Electronics, Movie-Book, and a three-domain Douban dataset) demonstrating consistent outperformance over state-of-the-art CDSR counterparts.

Experimental Results

Performance: CoRCi consistently outperforms all baselines (including ST-SR, DT-SR, MT-SR, and DT-CDSR models like ABXI, DREAM, and C2DSR) across all metrics (HR@5, HR@10, NDCG@10, MRR) on all datasets.
Statistical Significance: Improvements are statistically significant ( $p < 0.01$ ) in both domains of all datasets.
Ablation Studies:
- Replacing the SAE-CR hybrid with a single shared encoder or fully decoupled encoders results in performance degradation, validating the proposed architecture.
- Removing the Cross-Attention Transferors (CAT) or the stop-gradient operation significantly lowers performance, confirming their necessity for sequence-level fusion and maintaining coherence.
- Using standard InfoNCE instead of FocalNCE degrades performance, particularly in hard domains, highlighting the importance of suppressing intra-domain bias.
Generalizability: CoRCi maintains superior performance in a three-domain, non-overlapping user setting (Douban Movie-Book-Music), demonstrating robustness in complex, real-world scenarios.

Significance and Claims
The paper claims that CoRCi addresses the fundamental flaw in existing CDSR methods where per-domain loss aggregation and separate encoders disrupt the natural coherence of user interests across domains. By reconstructing mixed-domain representations from specific-domain encodings and utilizing a domain-agnostic loss with focal penalties, CoRCi effectively captures and preserves domain-invariant interests. The authors assert that this approach leads to more accurate user profiling and recommendation performance, particularly in scenarios where user interests evolve coherently across different product categories. The work suggests that preserving the "coherence" of interests is more critical than simply aggregating data, offering a new direction for modeling dynamic user preferences in cross-domain settings.

Technical Summary: CoRCi

More like this