What is URM
URM (Universal Recommendation Model) is a world knowledge large model launched by Alibaba Mama. Based on knowledge injection and information alignment, it combines the general knowledge of LLMs with professional knowledge in the e-commerce domain to address the problem of traditional LLMs performing poorly in e-commerce recommendation scenarios. URM introduces multimodal fused product representations and an efficient Sequence-In-Set-Out generation method to handle various recommendation tasks (such as multi-scenario recommendation, long-tail recommendation, etc.), significantly improving recommendation performance. URM has been launched in Alibaba Mama's display advertising scenarios, significantly improving merchants' advertising performance and consumers' shopping experience.
URM's main features
- Multi-task processing: Supports simultaneous processing of multiple recommendation tasks, such as multi-scenario recommendation, multi-objective recommendation, long-tail recommendation, discovery recommendation, etc.
- User interest understanding: Comprehensively understands users' interests and needs, providing recommendation results that better match users' current context.
- High-efficiency recommendation: Generates high-quality recommendation sets in a single forward pass, meeting the low-latency and high-QPS requirements of industrial-grade recommendation systems.
- Multimodal fusion: Combines product ID representations and semantic representations (such as text, images, etc.) to enhance product expressiveness and recommendation accuracy.
- Zero-shot learning capability: Quickly adapts to new tasks and new scenarios, providing effective recommendations without large amounts of annotated data.
URM's technical principles
- Multimodal fusion representation：
- ID representation: Based on a distributed hash table, product IDs are mapped to unique embedding vectors to capture collaborative information between products.
- Semantic representation: Text encoders and image encoders convert product text descriptions and image information into semantic embedding vectors.
- Fusion mechanism: The MLP layer aligns and fuses ID representations and semantic representations to generate multimodal product embedding vectors, enhancing product expressiveness.
- Sequence-In-Set-Out generation method: The input sequence consists of product IDs in user behavior, text tokens in task prompts, and special query tokens (such as UM, LM). The output mechanism generates user representations and text output separately based on the user modeling head (hUM) and the language model head (hLM). Increasing the number of UM tokens enables URM to generate multiple user representations in parallel in a single forward pass, improving recall metrics.
- Task definition and prompt engineering: Defines different recommendation tasks based on text, injects product IDs as special tokens into text descriptions, and achieves efficient expression of user behavior sequences. Designs different prompt templates and adjusts the distribution of recommendation results, allowing the model to dynamically adjust recommendation strategies according to different task requirements and user behavior.
- Training method: Combines the noise contrastive estimation (NCE) loss of the product recommendation task and the negative log-likelihood loss of the text generation task. Based on full-parameter supervised fine-tuning (SFT), only the original representations of products are frozen, preserving the LLM's pretrained knowledge.
- Asynchronous inference pipeline: To meet low-latency and high-QPS requirements, an asynchronous inference pipeline for dynamic capture of user behavior is designed, persisting inference results for the online recall stage to read. Based on multiple instances deployed in the same container, the concurrent QPS of URM inference is increased by 200%.
URM's project address
- arXiv technical paper：https://arxiv.org/pdf/2502.03041
URM's application scenarios
- Multi-scenario recommendations: Provide personalized recommendations based on different pages (such as the homepage and details page).
- Multi-objective recommendations: Comprehensively analyze user behaviors such as clicks and purchases to predict products of interest.
- Long-tail product recommendations: Uncover the potential value of low-frequency products and enrich user choices.
- Discovery recommendations: Recommend new categories of products that users have not encountered but may be interested in.
- Search-enhanced recommendations: Combine query intent and historical behavior to improve the relevance of search results.