LLM4Rec: Exploration and Application of Large Language Models in Recommendation Systems

This article presents a technical talk by Wang Yichao from Huawei Noah's Ark Lab on the exploration and application of Large Language Models (LLMs) in recommendation systems. The content is organized into three main sections: background and problems, LLM4Rec exploration and application, and challenges and prospects.

Background: Traditional recommendation systems are relatively closed systems that rely on user interaction behaviors (clicks, views) and item features (metadata like year, category, actors). They operate within specific application scenarios using logged data. In contrast, LLMs trained on internet-scale corpora possess rich world knowledge and logical reasoning capabilities that can complement traditional recommendation systems. LLMs have two key abilities that recommendation systems lack: (1) rich factual and commonsense knowledge providing deep details about items beyond the recommendation corpus, and (2) commonsense reasoning capabilities for analyzing item relationships and user behaviors.

LLM4Rec Exploration - KAR Project: The KAR (Knowledge-augmented Recommendation) project addresses the challenge of effectively extracting and storing LLM knowledge for traditional models. The approach decomposes knowledge generation into multiple sub-tasks using factorization. For movie recommendations, key factors like genre, director, actors, and awards are identified and integrated into prompt templates. The process involves three stages: (1) Knowledge Generation - using instruction templates to generate logical reasoning knowledge about user preferences and item factual information; (2) Knowledge Adaptation - converting generated knowledge into dense vectors using a multi-expert network to avoid high-dimensional information overwhelming the system; (3) Knowledge Utilization - integrating the adapted knowledge into traditional recommendation models for inference. Experimental results show significant AUC improvements, validating the effectiveness of open-domain knowledge in recommendation. The solution has been deployed in Huawei's app store and Huawei Music, significantly improving recommendation performance.

LLM4Rec Exploration - Uni-CTR Project: The Uni-CTR project focuses on cross-domain recommendation using LLMs as a multi-scenario recommendation foundation. Traditional multi-scenario recommendation faces challenges including dominant main scenarios affecting others, insufficient semantic information utilization, and significant impact when adding or removing scenarios. The proposed approach converts tabular data into natural language descriptions using prompt templates, then processes them through LLMs like SharedBert. Leader networks are introduced every few transformer layers to extract scenario-specific information, while the general network learns cross-scenario shared information. The solution also enables zero-shot cold start capability by leveraging LLM outputs to predict new scenarios. Experiments on Amazon Review datasets (Fashion, Music Instruments, Gift Cards) show significant performance improvements across all scenarios, especially in Gift Cards.

Challenges and Prospects: Current challenges include: (1) joint modeling of collaborative and semantic signals; (2) input strategy challenges including user profiling optimization, long text processing, and ID encoding fusion; (3) effectively integrating dynamic data and improving model inference efficiency for industrial applications. Future directions span three levels: Data - enabling recommendation systems with world knowledge and logical reasoning, moving from closed to open systems; Model - transitioning from discriminative to generative models with potential for directly generating user-interested content; Process - potentially replacing traditional multi-stage recommendation workflows with unified end-to-end models.