English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5) - RecSys 2022

Forum topic · 小凯 · 2026-07-05

Summary

This forum post discusses the RecSys 2022 paper "Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5)", which proposes recasting recommendation tasks as natural language processing problems. Under the P5 paradigm, items and user interactions are serialized into natural language, and a single pretrained language model is personalized with task-specific prompts to perform diverse recommendation objectives such as rating prediction, sequential recommendation, explanation generation, review summarization, and direct recommendation — all within one unified model instead of separate task-specific architectures. The post situates the work within sequential recommendation research, outlines the general methodology (problem formalization, prompt design, model training, and inference pipelines), and discusses evaluation practices, limitations, and engineering considerations for deploying LLM-based recommenders, including latency, cost, safety, and evaluation reliability. It also cross-references related sequential-recommendation work such as BERT4Rec, EAGER, Mamba4Rec, and follow-up research on item ID indexing for recommendation foundation models. Quantitative results should be verified against the original PDF via the ACM DL link (DOI: 10.1145/3523227.3546767).

Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5) — RecSys 2022

Source: ACM Digital Library — DOI 10.1145/3523227.3546767 Category: Sequential Recommendation

Overview

This RecSys 2022 paper introduces P5 (Pretrain, Personalized Prompt, and Predict Paradigm), which treats recommendation as a language processing problem. Instead of building separate models per task, P5 unifies recommendation objectives as natural language tasks handled by a single pretrained language model with personalized prompts.

Background and Motivation

Traditional recommendation pipelines separate retrieval, ranking, and generation, which makes it hard to support natural language interaction, multi-hop reasoning, and unified task handling. P5 responds to this by reformulating recommendation data (users, items, ratings, sequences) as natural language sequences, so a single model can be evaluated across multiple recommendation tasks under one framework.

Core Contributions

  • A unified perspective: recommendation tasks (rating prediction, sequential recommendation, explanation, review summarization, direct recommendation) are cast as text-to-text problems.
  • A prompt-based personalization scheme in the same spirit as prompting in NLP, enabling one pretrained model to serve many recommendation objectives.
  • A unified evaluation protocol across tasks, lowering the barrier for comparing recommendation approaches within a single framework.
  • Method Sketch (generalized)

    1. Input & representation: user histories, item metadata, and interactions are serialized into natural language prompts. 2. Core model: a pretrained language model is fine-tuned over a collection of personalized prompts. 3. Learning strategy: multi-task training over prompt variants. 4. Inference: task-appropriate prompts at prediction time, producing textual outputs that map back to recommendation decisions or explanations.

    Evaluation Notes

    Evaluation in this line of work typically spans recommendation accuracy (e.g., Recall@K, nDCG@K, Hit@K), explanation quality, and cross-task consistency. Exact numerical results should be taken from the original paper's tables rather than secondary summaries.

    Key Insights for Search / Rec / Personalization

    1. Architecture: unified prompt-based models reduce per-task engineering, while agentic paradigms are making retrieval strategies themselves learnable. 2. Data: high-quality instruction data and interaction logs are critical; synthetic data must guard against leakage and distribution shift. 3. Evaluation: gaps between offline metrics and online satisfaction persist; LLM-as-judge should be cross-validated with human evaluation. 4. Productization: latency, cost, interpretability, and safety are hard constraints for deployment.

    Limitations and Future Directions

    Potential limitations include experiment scale constrained by compute, benchmarks that may not match real user distributions, English-centric data limiting cross-lingual generalization, and safety risks of LLM-based generation (e.g., hallucinated explanations). Future work directions include more efficient test-time compute allocation, deeper integration with knowledge graphs and structured databases, and causal/fairness constraints for recommendation.

    Related Entries

  • BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations
  • EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration
  • Efficient On-Device Session-Based Recommendation (ACM TOIS)
  • How to Index Item IDs for Recommendation Foundation Models (P5, SIGIR)
  • LLMCDSR: Enhancing Cross-Domain Sequential Recommendation with Large Language Models
  • Mamba4Rec: Towards Efficient Sequential Recommendation with Selective State Space Models
  • References

  • Original paper: *Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5)*, RecSys 2022. https://dl.acm.org/doi/abs/10.1145/3523227.3546767

Tags

#recommendation-systems#llm#p5#prompt-learning#sequential-recommendation#recsys-2022#generative-recommendation

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178208985