Cross-Country Code-Mixing Feeds Big-Market Data to Small Markets in Generative Recommendation

Background: Parameter Sharing Without Data Interaction

Lazada operates across six countries with independently deployed systems; user interactions occur only within each country, so user IDs and item IDs are completely partitioned. Unlike classic cross-domain recommendation where shared users, items, or anchored entities transfer collaborative signals, cross-country e-commerce lacks any natural bridge. Generative recommendation (GR) maps each item to a discrete semantic token (SID) and autoregressively generates target SIDs from user history, enabling a unified embedding space and implicit knowledge sharing via model parameters (e.g., GenCDR, GMC). However, current GR approaches lack explicit data-level interaction : each country's interaction sequences remain domestic, so rich behavioral signals from large markets never directly reach small markets.

Inspiration from Code-Switching in LLMs

Research on large language models shows that cross-lingual ability stems from code-switching corpora — sentences mixing multiple languages (e.g., Chinese-English). This motivates constructing legitimate code-switching corpora in the multi-country recommendation setting to enable explicit cross-country data mixing.

Challenges in Constructing Cross-Country Mixed Corpora

Item-level similarity is ambiguous. Visual or textual similarity does not guarantee market interchangeability. Content similarity misses two critical signals: behavioral (co-click, co-purchase) and market dynamics (price band, audience, popularity). Two items that "look alike" may be a best-seller in one country and dead stock in another; using only content injects noise.

Context granularity is too coarse. Even if a pair of items seems reasonable in isolation, it may conflict with the current user intent. For example, inserting an Android tablet into an Apple-centric sequence is syntactically plausible but contextually wrong. Any offline "item-pair filter" cannot see this context.

CMRec: Cross-Country Code-Mixing with Dual Constraints and Context Awareness

2.1 Behavior + Content Unified Semantic Codebook

First, a multimodal LLM (Qwen2.5-VL) encodes item titles and images into content vectors. A dual-tower item-to-item (i2i) model injects behavioral information into these vectors. The combined vectors are then tokenized via RQ-VAE into token sequences. This codebook serves dual roles: tokenizer for GR and a shared semantic coordinate system for subsequent cross-country replacement.

2.2 Token-Level Mixing Under Dual Constraints

For each country's sequence, a proportion of historical items and the target item are replaced with nearest neighbors from other countries, synthesizing a "mixed-country sequence." Each replacement must satisfy two constraints:

Semantic content constraint: Hamming distance between token sequences ≤ threshold, ensuring semantic codes are sufficiently close.

Dynamic behavioral constraint: Side information (price, audience, popularity) is quantile-bucketed per country, then cosine similarity is computed to ensure market roles are also similar.

Only items meeting both constraints enter the neighbor set N(i) for replacement.

2.3 Context-Adaptive GR Training

The dual constraints guarantee content and behavioral similarity, but a mixed sample (x′, y′) derived from (x, y) may still be contextually illegal. A weight is defined to measure the contextual legality of the current replacement. The final training loss incorporates this weight to down-weight noisy mixed samples. (Exact formula omitted in source.)

Experimental Results

On public datasets and Alibaba industrial datasets, CMRec achieves state-of-the-art performance; both large and small countries improve, with larger gains for small countries.

Online A/B test at Lazada: ad revenue +1.77%, orders +2.64% .

References

[1] Recommender Systems with Generative Retrieval

[2] Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations

[3] Reg4Rec: Reasoning-Enhanced Generative Model for Large-Scale Recommendation Systems

[4] Sparse Meets Dense: Unified Generative Recommendations with Cascaded Sparse-Dense Representations

[5] Generative Recommendation Models: Progress and Directions

Paper: "Cross-Country Code-Mixing for Generative Recommendation" by Yuan Gao*, Hao Deng*, Haibo Xing, Yi Xu, Lingyu Mu, Jinxin Hu, Yu Zhang, Xiaoyi Zeng. Accepted at CIKM 2026 (Rome, Italy).

Code example

[2] Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations
[3] Reg4Rec: Reasoning-Enhanced Generative Model for Large-Scale Recommendation Systems
[4] Sparse Meets Dense: Unified Generative Recommendations with Cascaded Sparse-Dense Representations