关键词: 信息茧房, 算法推荐, 数字阅读, 阅读多样性, 群组轨迹模型, 状态转移

Abstract:

[Purpose/Significance] Algorithmic recommendation is widely used in digital reading services and shapes the range of content available to college students. Repeated exposure to similar topics may accompany a gradual concentration of reading interests. Research on information cocoons has examined selective exposure, platform design and user choice through surveys, experiments and behavioral records. Longitudinal evidence has concentrated mainly on news, music and video consumption. Less is known about how the topic diversity of college students' reading changes within an academic semester and how these changes differ across readers. This study examines the overall direction of change, distinct trajectories and transitions between diverse states. It provides evidence for academic libraries seeking to understand changes in reading breadth and develop reading promotion suited to different groups of students. [Method/Process] The study used de-identified reading-application logs covering 600 undergraduates at three universities in Inner Mongolia over 16 consecutive weeks. Participants used the same commercial reading application and represented different disciplines and years of study. Weekly reading diversity was measured using normalized topic entropy across 12 broad subject categories. Weeks containing fewer than six valid reading items were excluded from the valid observations. These weeks accounted for approximately 4.6 percent of student-week records and were interpolated between adjacent valid weeks for trajectory estimation. State transitions were calculated only from pairs in which both consecutive weeks were valid. A group-based trajectory model estimated patterns in the diversity sequences. The four-group linear specification was selected by considering group size, classification quality and the separation of trajectories. Three diversity states were defined using thresholds of 0.58 and 0.43, and their weekly transitions were summarized using 8 245 valid pairs. Alternative topic classifications, reading-volume treatments and state thresholds were examined in sensitivity analyses. A multinomial logistic regression assessed associations between trajectory membership, recommendation-entry use and active exploration, controlling for initial diversity, total reading volume, discipline and year of study. [Results/Conclusions] Mean reading diversity declined from 0.67 in the first week to 0.49 in the sixteenth week. The relative decline was 26.6 percent, calculated from unrounded weekly means. The interquartile range also shifted downward. Four trajectories were identified: diversity-stable, accounting for 27.7 percent of readers; gradually-narrowing, 34.8 percent; fast-cocooning, 25.7 percent; and low-lingering, 11.8 percent. The fast-cocooning group showed the steepest decline. The low-lingering group began with low diversity and experienced limited additional narrowing during the semester. Weekly staying probabilities were 0.79 for high diversity, 0.39 for medium diversity and 0.67 for low diversity. Direct transitions between the extreme states were uncommon. A likelihood-ratio test rejected a first-order Markov specification, so the transition matrix was used to describe adjacent-week changes without projecting long-term distributions. Greater recommendation-entry use and less active exploration were associated with membership in the narrowing or persistently low-diversity groups. These observational associations do not establish that recommendation algorithms caused the differences. Topic concentration may also reflect academic specialization, course requirements or stable interests. Academic libraries can monitor sustained changes in reading diversity with appropriate authorization and privacy protection, combine behavioral records with information about reading purposes, and offer broader subject choices to readers experiencing rapid narrowing. Active searching and exploration across topics merit attention in reading promotion. The findings are limited to active users of one application over one semester. Future research should include multiple reading settings, longer observation periods, platform-level recommendation records and experimental evaluations of reading interventions. Prospective classification should also be evaluated using later observations that are separate from the period used to construct early indicators.

Key words: information cocoon, algorithmic recommendation, digital reading, reading diversity, group-based trajectory model, state transition