[Paper Review] Understanding Longitudinal Dynamics of Recommender Systems with Agent-Based Modeling and Simulation

Gediminas Adomavičius, Dietmar Jannach+2 more|arXiv (Cornell University)|Aug 25, 2021

Consumer Market Behavior and Pricing29 references4 citations

TL;DR

This paper proposes agent-based modeling (ABM) as a simulation framework to study longitudinal dynamics in recommender systems, enabling the analysis of long-term effects such as popularity bias, trust erosion, and diversity loss. By simulating heterogeneous users, items, and a dynamic recommender engine over time, the approach reveals how algorithmic choices and feedback loops lead to emergent system-level phenomena that static experiments miss.

ABSTRACT

Today's research in recommender systems is largely based on experimental designs that are static in a sense that they do not consider potential longitudinal effects of providing recommendations to users. In reality, however, various important and interesting phenomena only emerge or become visible over time, e.g., when a recommender system continuously reinforces the popularity of already successful artists on a music streaming site or when recommendations that aim at profit maximization lead to a loss of consumer trust in the long run. In this paper, we discuss how Agent-Based Modeling and Simulation (ABM) techniques can be used to study such important longitudinal dynamics of recommender systems. To that purpose, we provide an overview of the ABM principles, outline a simulation framework for recommender systems based on the literature, and discuss various practical research questions that can be addressed with such an ABM-based simulation framework.

Motivation & Objective

• Address the gap in existing recommender system research, which focuses on static, offline evaluations and fails to capture long-term behavioral and systemic effects.
• Investigate longitudinal phenomena such as popularity reinforcement, homogenization of recommendations, and declining user trust that emerge only over time.
• Provide a simulation framework that models heterogeneous users, evolving item catalogs, and dynamic recommender engines to study system-level outcomes.
• Enable researchers to explore multi-stakeholder impacts—on consumers, content providers, and platforms—beyond short-term performance metrics.
• Support the evaluation of reinforcement learning-based recommendation strategies by creating controlled, longitudinal simulation environments.

Proposed method

• Model users as autonomous, heterogeneous agents with distinct preferences, consumption strategies, and feedback behaviors.
• Represent items as dynamic entities with lifespans, content descriptions, and availability timelines to simulate evolving catalogs.
• Implement the recommender engine as an agent that generates recommendations based on relevance prediction, ranking, and performance assessment.
• Simulate iterative interactions: users consume recommended items with probabilistic choice models, rate them, and optionally submit feedback.
• Update the rating database and recommender state after each interaction, while tracking active users and items based on modeled lifespans.
• Log system-level metrics (e.g., prediction accuracy, diversity, popularity concentration) across simulation steps to analyze longitudinal trends.

Experimental results

Research questions

• RQ1 How do different recommendation algorithms influence the long-term concentration of recommendations on popular items?
• RQ2 What are the long-term effects of profit-maximizing recommendation strategies on user trust and retention?
• RQ3 How do user feedback behaviors and consumption strategies shape the evolution of item popularity and system diversity?
• RQ4 To what extent do feedback loops reinforce existing popularity biases in recommender systems over time?
• RQ5 Can ABM-based simulations effectively model and benchmark reinforcement learning-based recommendation policies in a controlled, longitudinal setting?

Key findings

• Certain families of recommendation algorithms are more prone to popularity reinforcement and homogenization, leading to reduced diversity in long-term recommendations.
• Profit-maximizing strategies can lead to declining user trust over time due to repeated exposure to low-relevance or low-quality recommendations.
• Longitudinal simulations reveal that feedback loops amplify initial popularity imbalances, causing already popular items to dominate over time.
• The ABM framework successfully captures emergent system-level phenomena such as decreasing personalization and increasing concentration of recommendations.
• ABM provides a viable alternative to field experiments and offline evaluations for studying long-term dynamics that are difficult to observe in short-term studies.
• The framework can be extended to model multi-stakeholder environments, including social influences and content provider incentives, to study broader systemic impacts.