LLM-Driven Normative Recommendation Systems: From Stakeholder Demands to Agent Formalized Rules

This paper explores a normative recommendation system that uses large language models (LLMs) to extract stakeholder norms and formalize them into agent-executable rules. The core goal is to address the problem that traditional recommendation systems struggle to balance multi-party interests (such as listener preferences, artist exposure, platform business goals, etc.), with the DJ4ME music recommendation platform as a case study.

Traditional recommendation systems (such as collaborative filtering, content-based recommendation) often optimize a single objective and struggle to explicitly handle complex value trade-offs. Normative recommendation systems, on the other hand, focus on "morally/socially correct recommendations", stemming from concerns about the social impact of recommendation systems—algorithms not only reflect preferences but may also shape preferences, amplify biases, and affect creators' livelihoods. Their core assumption is that stakeholders' natural language concerns can be understood as "norms" (implicit rules), such as "emerging artists should have the opportunity to be discovered".

The project's technical route is divided into two phases:

1. Norm Extraction: Use LLMs to identify normative statements from stakeholder texts (interviews, policies, feedback), which requires distinguishing between descriptive (current state) and normative (ought-to-be) content.
2. Formalization Conversion: Map the extracted norms into logical representations (such as deontic logic, rule systems) so that recommendation agents can explicitly consider these norms when making decisions.

Taking DJ4ME as a case, the music recommendation scenario involves multi-party interest conflicts:

• Listeners: Discovering music they like
• Emerging artists: Exposure opportunities
• Established artists: Maintaining fan bases
• Platform: User retention and revenue
• Music industry: Copyright and fair distribution Over-optimizing short-term user satisfaction may lead to homogenization and harm emerging artists; mandatory promotion of new content may reduce user experience. The project attempts to use LLMs to extract actionable norms from these conflicts.

Advantages:

• Semantic understanding: Capture subtle language differences and identify implicit norms (without explicit deontic vocabulary).
• Context integration: Integrate fragmented demands to build a coherent normative picture.
• Scalability: Process large amounts of text at low cost, supporting large-scale applications. Challenges:
• Norm conflicts: Norms from different stakeholders may be contradictory, requiring mechanisms to resolve (priority, weighting, Pareto optimality).
• Formalization gap: Natural language norms are ambiguous; converting them into precise rules easily loses details or introduces errors.
• Verification difficulties: Lack of objective evaluation standards, making iterative optimization complex.

This project lies at the intersection of multiple fields:

• Value alignment: A core issue in AI safety, ensuring system behavior aligns with human values; this project is its application in the recommendation field.
• Multi-objective optimization: A long-standing topic in recommendation systems; the normative approach treats objective functions as constraint sets.
• Explainable AI: Formalized norms are naturally explainable, and norms can be cited as a basis for decisions.
• Computational social science: Using computational tools to study social norms, which can be extended to policy analysis, ethical review, and other fields.

This project proposes a new design paradigm: shifting from "what users will like" to "what we should recommend", recognizing that recommendation systems are socio-technical systems embedded in social value networks. Implications for practitioners:

• Explicitly identify and document the value assumptions behind recommendation strategies.
• Explain these assumptions in system documentation.
• Involve stakeholders in value trade-off decisions. Transparency itself is a normative commitment.

Limitations: The project is in the exploratory stage, with a limited case scope and unproven method generalizability; stricter evaluations are needed for LLM extraction reliability, formalization accuracy, and recommendation quality improvement. Future Directions:

• Develop robust norm conflict resolution mechanisms.
• Explore multi-modal inputs (norms in images, videos).
• Establish manual evaluation protocols for norm extraction quality.
• Apply the method to AI fields outside recommendation systems.