关键词: 智能推荐算法, 网络群体极化, 观点动力学, JA模型, 建模仿真

Abstract:

[Purpose/Significance] The spread of social media and intelligent recommendation algorithms has made online group polarization increasingly more covert and complex. By recommending content that matches users' historical preferences, algorithms create homogeneous information environments, producing "filter bubbles" and "echo chambers" that reduce tolerance for diverse views and push group opinions toward extremes. Such polarization undermines rational public discussion and may threaten social stability. Although existing research has revealed the influence of recommendation algorithms on opinion evolution, it has mainly focused on convergence speed and the number of opinion clusters, generally modelling recommendation as a binary contrast of presence versus absence. Thus, the quantitative relationship between core parameter values and polarization remains unexamined. Rather than assuming a unidirectional effect whereby algorithms reinforce homogeneous contact and aggravate polarization, this paper treats key algorithm parameters as adjustable variables in order to identify their bidirectional moderating effect. It explicitly embeds three platform-adjustable parameters－ recommendation scope, recommendation information volume and platform influence－into an opinion dynamics model containing assimilation, repulsion and neutrality rules, identifies non-monotonic parameter-polarization relationships, and proposes operable parameter intervals, thereby offering micro-level quantitative evidence for algorithmic governance. [Method/Process] Grounded in opinion dynamics theory, the study builds on the continuous-opinion paradigm and the Jager-Amblard model, which captures assimilation, repulsion and neutrality and fits real polarization better than the classical bounded-confidence model. Two improvements have been made. First, one-to-one interaction becomes one-to-many: each time a node interacts synchronously with all its neighbors and with all non-neighbor information recommended by the platform, using their mean opinion as the interaction benchmark. Second, the recommendation algorithm is embedded: the fixed convergence coefficient becomes platform influence, and two parameters are added－recommendation scope, the opinion-difference threshold that filters recommended content, and recommendation information volume, the number of non-neighbor nodes pushed per step. Simulations run on a Barabási-Albert scale-free network, reproducing the structural bias of social platforms toward high-influence nodes. Polarization rate, the proportion of nodes whose absolute opinion exceeds an extreme threshold, is the core metric; four assimilation-repulsion threshold combinations represent contexts ranging from highly contentious to high-consensus issues. [Results/Conclusions] Recommendation systematically lowers polarization, most markedly in extreme settings, where the polarization rate falls by about three quarters, though it may also create moderate but mutually isolated clusters resembling "information cocoons". Recommendation scope has a U-shaped effect: narrow scope deepens polarization, moderate scope suppresses it, and excessive scope causes rebound as moderate nodes converge while extreme nodes are driven further out. Recommendation volume shows diminishing marginal returns, and platform influence is non-linear－too low influence yields disorder without consensus, moderate influence fosters stable consensus, and high influence amplifies extreme interactions and triggers rebound. A governance framework is therefore proposed, combining a dual-channel platform architecture, parameter optimization anchored to these optimal intervals, dispute-sensitive dynamic regulation, and stress testing with parameter filing. Limitations include a static network, homogeneous users and simulation-only validation. Our future work will address dynamic structures, user heterogeneity, multi-issue coupling and empirical calibration.

Key words: intelligent recommendation algorithm, online group polarization, opinion dynamics, JA model, modeling and simulation