Introduction
In modern e-commerce and content platforms, search and recommendation systems are the two primary gateways for user discovery. While these systems often coexist on the same platform and serve the same users, they are traditionally treated as separate silos. This separation ignores a fundamental truth: user behavior in both scenarios is driven by a combination of stable, long-term preferences and immediate, context-specific needs.
This paper introduces UDITSR (Unified Dual-Intents Translation for joint modeling of Search and Recommendation), a framework designed to bridge this gap. By modeling user intent as a combination of "inherent" and "demand" factors and using search data to supervise recommendation logic, UDITSR achieves superior performance in both domains.
The Dual-Intent Challenge
User interactions are generally driven by two distinct types of intent:
- Inherent Intent: These are stable, long-term preferences, such as a user’s consistent liking for spicy food or high-end electronics.
- Demand Intent: These are fleeting, short-term needs, such as a sudden craving for a sandwich or the seasonal need for a winter coat.
In a search scenario, demand intent is explicit; the user provides a query that describes exactly what they want. In a recommendation scenario, however, demand intent is implicit and must be inferred from sparse interaction history. Previous joint models often failed to capture the dynamic nature of demand or the mathematical relationship between these two intent types and the final item selected.
Methodology: The UDITSR Framework
UDITSR addresses these challenges through three core technical innovations:
1. Search-Supervised Demand Intent Generator
To solve the problem of "guessing" implicit demand in recommendations, UDITSR utilizes real search queries from the same user as supervision. The model is trained to generate a "demand vector" for recommendation tasks that mimics the semantic information found in search queries. Essentially, the system learns to "hallucinate" the query a user would have typed if they were searching instead of browsing.
2. Dual-Intent Translation Propagation
Drawing inspiration from Knowledge Graph embedding techniques (such as TransE), UDITSR treats the relationship between intents and items as a geometric translation. The core logic follows the formula:
$$\text{Inherent Intent} + \text{Demand Intent} \approx \text{Interactive Item}$$
This relationship is embedded within a Graph Neural Network (GNN) structure. By propagating these "translations" across the user-item interaction graph, the model learns to update user and item representations in a way that respects both long-term tastes and immediate needs.
3. Intent Translation Contrastive Learning
To ensure the embedding space remains organized, the authors employ a contrastive loss function. This forces the combined vector (Inherent + Demand) to be physically close to the actual item the user interacted with, while simultaneously pushing it away from items the user ignored. This geometric alignment ensures that the "translation" logic holds up across millions of potential items.
Experimental Results
Offline Evaluation
The authors evaluated UDITSR using two proprietary datasets from Meituan, a leading Chinese takeaway platform. These datasets (MT-Small and MT-Large) contain millions of interactions, including both explicit search queries and implicit recommendation clicks.
UDITSR consistently outperformed state-of-the-art baselines, including LightGCN and DIN. On the MT-Small dataset, the model achieved an average improvement of 6.22% in search tasks and 3.06% in recommendation tasks, measured by Hit@5 and NDCG.
Online Deployment
Beyond offline metrics, UDITSR was deployed in a month-long A/B test on the Meituan Waimai (takeaway) home page. The real-world impact was significant:
- Gross Merchandise Value (GMV): Increased by 1.46%.
- Click-Through Rate (CTR): Increased by 0.77%.
In the context of high-volume e-commerce, these percentage gains represent substantial increases in revenue and user engagement.
Conclusion and Future Work
UDITSR demonstrates that search and recommendation are two sides of the same coin. By viewing demand intent as a "deviation" from a user's base personality and using a geometric translation framework, the model provides a more accurate representation of user behavior.
While highly effective, the model currently faces challenges with "cold start" scenarios—new users or items with no search history. Future research will focus on incorporating temporal dynamics (the time elapsed between interactions) and expanding the framework to include other discovery scenarios like social feeds and live-streaming commerce.
Summary updated