Agent4Rec

Creates 1,000 autonomous agents initialized from MovieLens-1M user data, each embodying distinct social traits (conformity, activity, diversity preferences) and personalized movie preferences. Agents use LLM-based decision-making to generate realistic reactions to recommendations, retrieving contextual memories of past interactions and synthesizing responses that reflect individual behavioral patterns rather than deterministic algorithms.

Each agent maintains a persistent memory system that stores past interactions (watched movies, ratings, evaluations, exits) and retrieves relevant memories when deciding how to respond to new recommendations. The memory system uses semantic or temporal retrieval to surface contextually relevant past experiences, which the LLM then incorporates into its reasoning to generate consistent, history-aware decisions rather than stateless responses.

Provides a pluggable architecture for integrating multiple recommendation algorithms (Matrix Factorization, MultVAE, LightGCN, baseline models) into a unified simulation framework. The Arena component orchestrates the flow of user-item interactions through selected recommender models, collecting predictions and passing them to agents for evaluation. Models are loaded from configuration, trained or pre-trained, and called in a standardized way regardless of underlying implementation.

Simulates realistic user-recommendation interactions by presenting items in pages (multiple recommendations per round) and allowing agents to take diverse actions: watch, rate, evaluate, exit, or respond to interviews. Each action is generated by the LLM based on the agent's persona, memory, and the presented recommendations, creating a multi-step interaction loop that mirrors how users browse and interact with recommendation interfaces.

Initializes 1,000 agents by extracting user personas from MovieLens-1M dataset, deriving each agent's movie preferences, social traits (conformity, activity level, diversity preferences), and demographic characteristics from real user rating patterns. The initialization process maps historical user behavior to agent attributes, enabling agents to exhibit preferences grounded in actual user data rather than synthetic or random distributions.

Computes standard recommendation evaluation metrics (click-through rate, conversion, diversity, fairness) from agent interaction logs and performs causal analysis to understand how recommendation algorithm choices affect user behavior. The evaluation framework aggregates agent actions across the simulation, calculates metrics per model, and enables comparative analysis of how different recommenders influence agent engagement and satisfaction.

Provides a configuration-based system for defining and running recommendation simulation experiments, specifying which recommender models to evaluate, agent parameters, interaction settings, and evaluation metrics. The Arena component reads configuration files, initializes the simulation environment, orchestrates the interaction loop across all agents and models, and collects results in a structured format for analysis.

Integrates advertisement or sponsored items into the recommendation simulation, allowing evaluation of how agents respond to ads mixed with organic recommendations. The system can inject sponsored items into recommendation pages and measure agent engagement (clicks, watches, ratings) with ads versus organic items, enabling analysis of ad effectiveness and potential bias in recommendation algorithms.

Supports parallel execution of agent interactions across multiple processes or machines, enabling simulation of 1,000+ agents at scale. The Arena component can distribute agent-model interactions across available compute resources, collecting results from parallel workers and aggregating them into final metrics. This architecture allows simulations to complete in reasonable time despite the computational cost of LLM-based decision-making per agent.