Studying the Effects of Cognitive Biases in the Recommendation Interaction Feedback Loop Via Simulation

This study investigates the emergence and reinforcement of cognitive biases, specifically confirmation bias, within the dynamic interplay of user-recommender system interactions. We address two primary research questions:
- RQ1: Can confirmation bias be inferred from historical user interactions using tag-weighted modifications to a multinomial logit choice model?
- RQ2: How do different choice models and recommendation strategies – such as user weighting and rule-based item selection – affect the dynamics of the feedback loop?
Our objective is to identify and characterize measurable patterns indicative of bias rein- forcement through a robust simulation framework.
Utilizing the LFM-1b dataset, our methodology employs both standard and modified multinomial logit choice models alongside distinct recommendation strategies, including Repetition-Tolerant Exposure and Tag-Based Filtering. We evaluate the behaviors of MostPop, ItemKNN, and ALS models, quantifying confirmation bias via the odds ratio and analyzing temporal shifts in user preferences and recommendation outputs with the Mann–Kendall test.
Our simulations demonstrate moderate evidence of preference reinforcement, particularly in the absence of explicit diversity-promoting strategies. We observe how varying choice models and recommendation approaches impact feedback loop dynamics, revealing subtle yet significant shifts in user tag preferences and recommended item compositions that indicate bias perpetuation. These findings underscore the critical importance of understanding user-recommender system interactions in the context of cognitive biases and contribute to the development of more transparent and equitable algorithmic design.