A Simulation Framework for Studying Systemic Effects of Feedback Loops in Recommender Systems

TL;DR

This paper presents a flexible simulation framework to study systemic effects of recommender feedback loops using implicit purchases and epoch-based retraining on Amazon data. It shows a robust trade-off: increasing recommender influence expands individual purchase diversity while reducing collective diversity and concentrating purchases on a small set of items, with some models also increasing user homogenization. The framework supports multiple recommendation algorithms and epoch-based retraining to disentangle these effects across adoption-rate scenarios. The findings underscore the need for designs that balance personalization with long-term ecosystem diversity and resilience.

Abstract

Recommender systems continuously interact with users, creating feedback loops that shape both individual behavior and collective market dynamics. This paper introduces a simulation framework to model these loops in online retail environments, where recommenders are periodically retrained on evolving user-item interactions. Using the Amazon e-Commerce dataset, we analyze how different recommendation algorithms influence diversity, purchase concentration, and user homogenization over time. Results reveal a systematic trade-off: while the feedback loop increases individual diversity, it simultaneously reduces collective diversity and concentrates demand on a few popular items. Moreover, for some recommender systems, the feedback loop increases user homogenization over time, making user purchase profiles increasingly similar. These findings underscore the need for recommender designs that balance personalization with long-term diversity.

Figures (4)

• Figure 1: Individual and collective diversity. The evolution of individual (a) and collective (b) Gini coefficients as a function of adoption rate. Below, individual Gini coefficient disaggregated by user engagement segments for two representative models: NeuMF (c) and LightGCN (d).
• Figure 2: Purchase redistribution for $\eta=1$. Frequency-rank curves for the (a) purchase volume and (b) item popularity. Curves show the initial state (training set, grey) and the final state after 24 simulation epochs for NeuMF (green) and LightGCN (red).
• Figure 3: Co-purchase networks before and after the feedback loop unfolds (LightGCN, $\eta = 1$, 24 epochs). Each network includes a sample of items from different popularity levels. Nodes represent items (sized by total purchases), and edges connect items with shared buyers (thicker edges mean more shared purchasers).
• Figure 4: User homogenization. (a) User similarity as a function of adoption rate ($\eta$): LightGCN and MultiVAE strongly increase homogenization, BPR and Spectral CF show moderate growth, while NeuMF and Item-KNN reduce it at high adoption. (b) For ($\eta=0.8$), NeuMF remains stable, while LightGCN steadily amplifies user similarity over time.