RecAI: Reshaping Recommendation System Interaction and Interpretability with Large Language Models

RecAI explores integrating large language models into traditional recommendation systems, aiming to solve the problems of lack of transparency and user control in traditional systems. Its core directions are to enable users to shift from passively receiving recommendations to actively participating in the recommendation process through natural language interaction, interpretable recommendations, and fine-grained control, thus reshaping the human-computer interaction model of recommendation systems.

Recommendation systems have permeated all aspects of digital life. Models like collaborative filtering and deep learning have improved precision, but there is a long-neglected black box problem: users do not understand the recommendation logic and cannot effectively intervene. This opacity prevents users from correcting wrong inferences, expressing immediate preferences—they can only accept recommendations entirely or turn them off, lacking a middle ground for fine control.

The core design concepts of RecAI include three key terms:

1. Interactivity: Introduce natural language dialogue capabilities, allowing users to actively express preferences (e.g., "Recommend movies for relaxing on weekends") and break the limitation of relying on historical behavior;
2. Interpretability: Use natural language to explain recommendation reasons (e.g., linking to works users like or director styles) to enhance user trust;
3. Controllability: Support fine-grained preference adjustment (e.g., "Reduce the proportion of action movies and increase documentaries") to let users dominate the recommendation process.

RecAI adopts a hybrid architecture where LLM collaborates with traditional recommendation components:

1. Knowledge-Enhanced Recommendation Generation: Combine external knowledge bases (e.g., movie metadata) through knowledge plugins to make up for LLM's lack of domain-specific information;
2. Multi-turn Dialogue State Management: Track contextual preferences and understand complex needs (e.g., "Comedies but newer ones");
3. Controllable Generation Mechanism: Filter and sort based on users' explicit conditions to ensure recommendations align with control intentions.

RecAI is applicable to multiple scenarios:

• Content Streaming: Help discover interesting content and explain the reasons;
• E-commerce Shopping Guidance: Support complex shopping needs (e.g., "Lightweight outdoor jackets for spring, within a budget of 500 yuan") and explain recommendations;
• News and Information: Respond to users' immediate information needs (e.g., "Progress in the field of artificial intelligence");
• Learning Resources: Recommend materials based on goals and levels and explain the reasons.

RecAI faces challenges in deployment:

1. Inference Cost: LLM has high inference costs, requiring a balance with real-time recommendation response speed;
2. Hallucination Problem: LLM may generate false recommendation reasons, requiring fact-checking mechanisms;
3. User Habits: The interactive mode requires users to invest more effort, so a balance between feature richness and convenience is needed.

RecAI represents the direction of recommendation systems moving from algorithm-driven black boxes to transparent models of human-machine collaboration. The interaction and interpretation capabilities brought by LLM enable users to become participants and leaders in the recommendation process. As technology matures and costs decrease, future recommendation systems will be more intelligent and transparent, respecting users' right to autonomous choice.