URM - A World Knowledge Model launched by Alibaba's Alimama

What is URM?

URM (Universal Recommendation Model) is a global knowledge model launched by Alimama. Based on knowledge injection and information alignment, it combines the general knowledge of LLM with the professional knowledge of the e-commerce field to solve the problem of poor performance of traditional LLM in e-commerce recommendation scenarios. URM introduces multimodal fusion of product representation and 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 Alimama's display advertising scenario, significantly improving the performance of merchants' advertising and the shopping experience of consumers.

Main functions of URM

• MultitaskingIt supports simultaneous processing of multiple recommendation tasks, such as multi-scenario recommendation, multi-objective recommendation, long-tail recommendation, and discovery-based recommendation.
• User Interest UnderstandingTo fully understand users' interests and needs, and provide recommendation results that are more in line with the user's current context.
• High-efficiency recommendationGenerates a high-quality recommendation set in a single forward propagation, meeting the low latency and high QPS requirements of industrial-grade recommendation systems.
• Multimodal fusionBy combining product ID representations and semantic representations (such as text and images), we can improve the expressive power of products and the accuracy of recommendations.
• Zero-shot learning abilityIt can quickly adapt to new tasks and scenarios and provide effective recommendations even without a large amount of labeled data.

URM's technical principles

• Multimodal fusion characterization:
  • ID representationBased on a distributed hash table, product IDs are mapped to unique embedding vectors to capture collaborative information between products.
  • Semantic representationText encoders and image encoders convert text descriptions and image information of products into semantic embedding vectors.
  • Integration mechanismThe MLP layer aligns and fuses ID representation and semantic representation to generate multimodal product embedding vectors, thereby enhancing the expressive power of products.
• Sequence-In-Set-Out generation methodThe input sequence consists of the product ID from user behavior, the text token from task prompts, and special query characters (such as UM and LM). The output mechanism generates user representations and text outputs based on the user modeling head (hUM) and the language model head (hLM), respectively. Increasing the number of UM tokens allows URM to generate multiple user representations in parallel during a single forward propagation, improving recall metrics.
• Task definition and prompting engineeringThis approach defines different recommendation tasks based on text, injecting product IDs as special tokens into the text descriptions to efficiently represent user behavior sequences. Different prompt templates are designed to adjust the distribution of recommendation results, allowing the model to dynamically adjust its recommendation strategy according to different task requirements and user behavior.
• Training methodsThis approach combines the Noise Contrast Estimation (NCE) loss from product recommendation tasks with the Negative Log-Likelihood loss from text generation tasks. Supervised Fine-Tuning (SFT) based on complete parameters freezes only the original product representations, preserving the pre-trained knowledge of the LLM.
• Asynchronous inference linkTo meet the requirements of low latency and high QPS, an asynchronous inference pipeline for dynamically capturing user behavior is designed, and the inference results are persistently stored for retrieval during the online recall phase. By deploying multiple instances in the same container, the concurrent QPS of URM inference is increased by 200%.

URM project address

• arXiv technical paper:https://arxiv.org/pdf/2502.03041

URM application scenarios

• Multi-scenario recommendationProvide personalized recommendations based on different pages (such as the homepage and the details page).
• Multi-objective recommendation: By comprehensively analyzing user clicks, purchases, and other behaviors, we can predict products that users are interested in.
• Long-tail product recommendationsExplore the potential value of low-frequency products and enrich user choices.
• Discovery-based recommendationsRecommend new product categories that users may not have encountered before but might be interested in.
• Enhanced search recommendationsBy combining query intent and historical behavior, the relevance of search results can be improved.