An Unified Search and Recommendation Foundation Model for Cold-Start Scenario

Yuqi Gong, Xichen Ding, Yehui Su, Kaiming Shen, Zhongyi Liu, Guannan Zhang

Abstract

Traditional digital platforms often treat Search (S) and Recommendation (R) as distinct tasks, leading to modeling inefficiencies and a persistent "cold-start" problem for new items or services. This paper presents a unified S&R Multi-Domain Foundation Model developed by researchers at Ant Group. By integrating Large Language Models (LLMs) for semantic feature extraction and employing a novel Aspect Gating mechanism, the model effectively bridges the gap between search and recommendation. Deployed within the Alipay app, the foundation model demonstrates significant performance gains in cold-start scenarios and real-world online metrics.

1. Introduction

In modern e-commerce and service ecosystems, Search and Recommendation are the primary drivers of user discovery. Search is typically characterized by explicit user intent via queries, while Recommendation relies on implicit signals from browsing history. Historically, these systems have been developed in silos, relying heavily on unique identifiers (IDs) for users and items. This ID-centric approach fails when encountering new services or items that lack historical interaction data—the "cold-start" challenge.

This research proposes a paradigm shift: treating Search and Recommendation as two sides of the same coin within a unified foundation model. By anchoring the system in the universal semantic meaning of text rather than discrete IDs, the model can generalize across different domains and tasks.

2. Methodology

The S&R Multi-Domain Foundation Model is built upon three core technical innovations designed to facilitate cross-domain knowledge transfer and robust feature representation.

2.1 LLM-based Feature Extraction

To move beyond the limitations of ID-based features, the model utilizes pretrained LLMs (such as BERT or ChatGLM) to extract domain-invariant text features from search queries and item titles. Because the semantic meaning of a term (e.g., "financial services") remains constant whether it appears in a search bar or a recommendation feed, these LLM-derived embeddings provide a stable, shared representation space across all tasks.

2.2 Aspect Gating Fusion

Not all features are equally relevant in every scenario. In established domains with rich history, ID features are highly predictive; however, in cold-start domains, semantic text features are more critical. The model introduces an Aspect Gating mechanism that dynamically weights three distinct feature categories: ID-based features, LLM-extracted text features, and heterogeneous sparse features. This allows the model to adapt its focus based on the maturity and data density of the specific domain it is processing.

2.3 Domain Adaptive Multi-Task Learning (MTL)

To prevent "negative transfer"—where learning one task inadvertently degrades performance on another—the architecture employs a Domain Adaptive Layer. The researchers use Jensen-Shannon (JS) Divergence as a regularizer to align the distributions of different domains. This forces the model to map diverse inputs into a common vector space, ensuring that the learned representations are truly generalizable and foundational.

3. Experimental Evaluation

3.1 Datasets

The model was rigorously tested on seven large-scale industrial datasets from the Alipay mobile app. These datasets encompass a wide variety of tasks, including:

  • Search: Query-item relevance and content search ranking.
  • Recommendation: Background word query and content query recommendations.
  • Hybrid Tasks: Query suggestion Click-Through Rate (CTR) and service card recommendations.

The scale of these datasets, ranging from 0.76 million to 146 million samples, ensures the findings are applicable to high-traffic production environments.

3.2 Performance Results

The foundation model consistently outperformed established multi-task architectures such as MMoE and PLE. The most significant impact was observed in cold-start scenarios:

  • Cold-Start Gains: In specific cold-start tasks, the model achieved absolute Area Under the Curve (AUC) improvements of +0.0216 and +0.0279 over single-domain baselines.
  • Online Impact: During live A/B testing on the Alipay platform, the model delivered a +17.54% relative gain in Page View Click-Through Rate (PVCTR) for service card recommendations.

4. Discussion and Future Work

While the model demonstrates the power of unifying S&R tasks, the researchers identified several areas for further refinement. The use of large-scale LLMs (6B+ parameters) introduces computational complexity, which was managed by using them as offline feature extractors. Additionally, the study found that simple prompt engineering did not always improve performance, suggesting that the optimal way to "query" LLMs for recommendation-specific features remains an open research question.

Future iterations of the model may explore the integration of multi-modal features, such as image and video data, to further enhance the unified representation and provide even more robust performance in data-sparse environments.

5. Conclusion

The S&R Multi-Domain Foundation Model represents a significant step toward breaking down the silos between search and recommendation. By prioritizing semantic text features and employing adaptive gating, the system provides a scalable solution to the cold-start problem, as evidenced by its successful deployment and substantial performance improvements in a major industrial application.

Summary updated

BibTeX
@inproceedings{gong2023unified,
  author = {Yuqi Gong and Xichen Ding and Yehui Su and Kaiming Shen and Zhongyi Liu and Guannan Zhang},
  title = {{An Unified Search and Recommendation Foundation Model for Cold-Start Scenario}},
  booktitle = {ACM International Conference on Information and Knowledge Management},
  year = {2023},
  doi = {10.1145/3583780.3614657},
}
Figure