Short-form videos have become deeply embedded in college students' daily lives, raising substantial concerns regarding their high intensity of use and associated dependency risks. Short-form addiction tendency is closely linked to academic decline, psychological distress, and impaired social adjustment. The I-PACE model provides a comprehensive framework for understanding such addictive behaviors. However, existing studies have largely examined a limited set of variables within isolated theoretical perspectives, thus lacking a holistic analysis of the combined effects of multi-dimensional factors across the user-affect-cognition-execution (UACE) domains. Machine learning is adept at automatically discerning complex non-linear relationships among multiple variables, enabling more accurate individual-level risk prediction. This approach has been successfully applied in predictive research on issues such as depression, self-harm, suicide risk, and academic performance, demonstrating robust predictive utility. This study aims to integrate machine learning methodologies with the I-PACE theoretical framework. By incorporating multiple susceptibility factors spanning the four core modules, namely, user characteristics, affective states, cognitive processes, and executive functions, we developed and validated a predictive model for college students' short-form video addiction tendency and identified the core susceptibility factors contributing to it.

Guided by the Interaction of Person-Affect-Cognition-Execution (I-PACE) model, fourteen individual and contextual variables related to short-form video addiction tendency were selected. The selected variables comprised the four core domains of the framework: Person (P) factors, which included gender, age, neuroticism, boredom proneness, life stress, stress coping styles (both active and passive coping), and usage motivations (information-seeking, pastime, and entertainment); Affect (A) factors, namely emotional experiences during use, encompassing both positive and negative affective states; Cognition (C) factors, represented by cognitive bias; and Execution (E) factors, specifically inhibitory control. The survey was administered to 1,274 college students in Shanghai, Jiangxi, Anhui, and Guizhou. Five machine learning models were constructed, including Random Forest (RF), Logistic Regression (LR), Support Vector Machine (SVM), Decision Tree (DT), and K-Nearest Neighbors (KNN), and their predictive performance was compared.

Compared with the non-addiction-tendency group, the addiction-tendency group scored significantly higher on neuroticism, boredom proneness, perceived stress, passive coping style, pastime motivation, entertainment motivation, positive affect during use, and negative affect during use, while scoring significantly lower on active coping style, cognitive bias, and inhibitory control. The Random Forest model demonstrated the best performance, with a precision of 67.86%, a recall of 74.51%, an F1-score of 71.03%, and an area under the ROC curve (AUC) of 0.75. Feature importance analysis showed that the top five susceptibility factors were cognitive bias (21.52%), boredom proneness (19.92%), inhibitory control (16.88%), negative use-related emotions (11.72%), and pastime motivation (6.99%).

The Random Forest model constructed in this study can effectively identify college students at high risk of short-form video addiction tendency. Boredom proneness, inhibitory control, negative use-related emotions, and entertainment motivation were the most significant vulnerability factors for short-form video addiction tendency. Based on the I-PACE framework, the susceptibility factors for short-form video addiction tendency exhibit a hierarchical structure: at the Person level, boredom proneness and recreational motivation initiate use; at the Affect level, negative emotions reinforce dependency; at the Cognition level, cognitive bias undermines self-regulation; and at the Execution level, deficient inhibitory control leads to loss of control. These findings provide clear targets for screening and intervention.

Drawing on the I-PACE model, this study developed a short-form video addiction tendency prediction model using machine learning algorithms. The Random Forest model demonstrated the best performance. Variable importance analysis revealed that cognitive bias, boredom proneness, inhibitory control, negative use-related emotions, and entertainment motivation were identified as the top five susceptibility factors. The model can assist educators and clinicians in efficiently screening college students at high risk of short-form video addiction tendency. Furthermore, these key susceptibility factors may serve as targets for prevention and intervention in both educational and clinical settings, thereby helping to reduce the prevalence of short-form video addiction tendency. This, in turn, promotes the healthy physical and mental development of college students.

The random forest prediction model developed in this study can serve as a supplementary screening tool in university mental health screenings, helping institutions identify students at high risk of short-form video addiction early on. As college students' tendency toward short-form video addiction is closely associated with cognitive biases, boredom proneness, inhibitory control, negative use-related emotions, and recreational use motivation, universities should consider implementing targeted prevention strategies, including cognitive bias modification, boredom management, self-control enhancement, emotion regulation training, and support for alternative activities.