Abstract:

[Purpose/Significance] The use of short-video algorithmic recommendation services poses risks such as privacy breaches, information cocoons, cognitive manipulation, and degraded content quality. Existing studies mostly focus on risk types and consequences, lacking a systematic explanation of the generative mechanism of users' risk-coping behaviors, especially the internal logical relationships among behavioral factors. This study integrates the risk information seeking and processing (RISP) model and the protective action decision model (PADM) to deconstruct the generative mechanism of short-video users' risk-coping behaviors toward algorithmic recommendation services. The innovation lies in synthesizing cognitive, affective, and behavioral dimensions into a unified model, extending risk protection theories into the algorithmic context, and providing practical guidance for platform algorithm optimization and governance. [Method/Process] Using the critical incident technique (CIT), we collected data through semi-structured interviews and open-ended questionnaires from active short-video users, who had experienced at least one identifiable algorithm-related risk incident. Participants represented diverse demographic backgrounds, spanning different age groups, education levels, and usage frequencies to ensure broad representativeness. Thematic analysis was applied to code the data, examining behavioral elements and their interactions across four sequential stages: attention triggering, risk assessment, emotional response, and behavioral control. Cross-case comparisons were used to map interdependencies among these stages, leading to the development of a comprehensive generative mechanism model. [Results/Conclusions] The study reveals that risk-coping behaviors systematically unfold through four stages: attention triggering, risk assessment, emotional response, and behavioral control. These stages exhibit iterative relationships, with feedback loops also existing between emotions and risk reappraisal. Based on this framework, we propose corresponding management strategies for each stage: during attention triggering, algorithms should be optimized to align with users' cognitive capacities; during risk assessment, bidirectional feedback channels should be enhanced to encourage deeper user engagement; during emotional response, risk perception biases should be mitigated while harnessing the positive potential of emotional responses; and during behavioral control, users' algorithmic rights must be safeguarded, balancing technological power with user responsibility. This study offers an integrative explanation of the coupled dynamics among cognitive, emotional, and behavioral factors within algorithmic risk contexts. Limitations include inherent recall bias in qualitative CIT designs and limited generalizability. Future research should employ quantitative methods to validate the model empirically and test its causal pathways. This will strengthen the model's theoretical and practical foundations for algorithmic governance.

Key words: risk protection, short video users, algorithmic recommendation service, risk response behavior