Beyond Accuracy: Understanding and Modeling the Role of User Conformity in Recommender Systems

Recommender Systems (RSs), particularly Collaborative Filtering (CF) models, commonly employed in e-commerce domains, face challenges in adapting to sudden population-scale events, such as the COVID-19 pandemic and annual Black Friday shopping events. My initial study uncovers a phenomenon termed "population-scale concept drift" in user interaction behavior. During pandemic-like events, individuals exhibit imitative behavior, leading to abrupt shifts in their interactions with CF systems. Grounded in the theory of user needs, our simulation-based framework, the Need Evolution Simulator (NEST), utilizing model-agnostic reinforcement learning, investigates the impact of rapid drifts in user behavior on RSs. By simulating the evolution of human needs, we analyze macro-trends in population dynamics, providing insights into the influence of events on CF models. Experimental results reveal an initial impact on CF performance during the early stages of events, followed by an exacerbated population herding effect. This effect introduces a popularity bias, benefiting some users but compromising the overall user experience. To address this, we propose an adaptive ensemble method that optimally adapts algorithms to different event stages. In our second study, we comprehensively examine temporal dynamics in user behavior and their impact on recommendation systems under both pandemic-related contingencies and calm circumstances. Drawing on user need theory, we highlight the irrationality of individuals' decision-making, often influenced by peers in the same social community. Our community detection-based evaluation approach effectively identifies collaborative concept drifts among users, shedding light on the evolution of herds in RS under extreme outlier events and calm environments. Findings indicate that CF models may achieve high accuracy by recommending popular items during population-level herd behavior, limiting recommendation diversity. Traditional evaluation methods focusing solely on accuracy may not comprehensively measure RS performance. Additionally, our third work investigates the influence of user conformity behavior on RSs. Conventional RSs assume user behavior is solely driven by individual interests, overlooking the impact of peer influence and resulting conformity behavior. Solutions that indiscriminately eliminate such bias may depersonalize recommendations. We propose the Temporal Conformity-aware Hawkes Network (TCHN) model, based on Hawkes processes, to identify two forms of conformity behavior: informational conformity and normative conformity. TCHN disentangles user interest and conformity in a personalized manner, incorporating attention-based methods to model stable and volatile dynamics. Experiments on real-world datasets demonstrate varying conformity scales among users, with TCHN exhibiting advantages in accuracy and diversity in recommendations.