Predictive Analysis of College Students' Creative Self-Efficacy Based on Decision Tree Modeling

Summary

The study examines the use of a modified C5.0 decision tree model (DTM) to predict creative self-efficacy (CSE) among college students. The research aims to determine the effectiveness of DTM in predicting CSE and identify the variables that predict it.

Abstract

Creativity is a focal point in current psychological research, domestically and internationally. An individual's self-awareness of their creativity is vital for determining their creative potential, helping them understand their strengths and weaknesses, and enhancing their self-creativity. There are few studies on college students' self-perception of creativity and even fewer on using the decision tree (DT) to predict such self-perception. This study aimed to assess the effectiveness of using a modified C5.0 decision tree model (DTM) in predicting creative self-efficacy (CSE) among college students and to identify which variables could serve as predictors of CSE. In response to these questions, 607 college students in Sichuan Province, China, filled out a Strengths and Difficulties Questionnaire (SDQ) containing five sections: psychological resilience scale, academic self-efficacy scale, psychological sense of school membership, innovative behavior scale, and CSE scale. The method described involves systematically providing the SDQ, scoring the results, and assigning codes to the answers of 60 students. Using the SDQ sub-scales, a DTM called C5.0 was built to predict the types of behavioral risk. The DTM was trained and cross-validated, showing a 10-fold improvement, and its performance was evaluated by accuracy and F1-score. The results of the revised C5.0 DTM revealed that the significant predictive factors of CSE, in decreasing order of importance, are psychological trust, psychological sense of school membership, academic self-efficacy, and innovative behavior. By proving the DTM's effectiveness in predicting college students' CSE, this research developed the link between basic learning and CSE, increased the interpretation of CSE, and provided research support for enhancing college students' creativity and CSE.

Introduction

Creativity is regarded as an essential motivator for the socio-economic development of several countries and has progressively become a primary factor in a nation's overall strength. Nevertheless, research indicates that Chinese Higher Education (HE) institutions still prioritize students' theoretical knowledge, neglecting the cultivation of analytical opinion skills, autonomous problem-solving skills, and innovative skills. Consequently, several students possess strong theoretical knowledge but struggle to analyze and address question¹. The dual impacts of personal and social environments drive this phenomenon, motivating college students to prioritize learning and external social competition over their emotional needs and personal development. A significant risk has emerged in the form of Artificial Intelligence (AI), on which modern society has become overly dependent, to the point of diminishing individual opinion and innovative skills. This is the result of a limitation or regression in personal and social development^2,3.

Therefore, innovation has emerged as a Hot Topic in current domestic and foreign psychological research, with many studies signifying that Creative Self-Efficacy (CSE) positively impacts creativity. Simultaneously, research indicates that higher CSE among students correlates with a more positive outlook on life and increased confidence in their skills when facing challenging tasks^4,5. Indeed, low levels of creativity can adversely affect all features of life, including a competitive problem in one's career. Studies have shown that the primary cause of poor creativity is a lack of CSE. As a significant purpose of creativity, CSE is defined as an individual's belief in their capacity to make creative thoughts, participate in creative behaviors, and achieve creative results. In teams, CSE can increase individual creativity in knowledge sharing and collaborative efforts⁶. High CSE can also inspire new Innovative Behavior (IB) in the presence of optimistic emotions⁷.

As advised by changing psychology, around the ages of 10 to 14, students' skills may lead to changes in their emotions and behavior because adulthood is happening⁸, and they are dealing more with peers and the stress that comes with it. Because the SDQ has been confirmed to be trusted and psychometrically sound and covers only 25 items⁹, it was preferred for this research. It looks at five types of behavior -- emotions, actions at university, activity level, relationships at college, and helpfulness -- to illustrate the general state of a student's mental health. According to existing works^4,5, they found that using the SDQ in China is more effective than using behavioral tasks alone in identifying the early presence of psychological and behavioral challenges in pupils. It displays why investigation is essential and works well in several university environments.

Given that college students are the core force and leading contributors to future society, their CSE holds particular importance. Research has revealed that students' CSE is directly related to positive emotions and control direction⁸. Additionally, CSE improves students' behaviors and performance in extracurricular activities⁹. Research further indicates that students' CSE enhances their social skills and Psychological Resilience (PR), enabling them to adapt to challenging environments¹⁰. Students' PR positively impacts their flexibility, coping abilities, intellectual skills, and emotional inhibition¹¹. Therefore, motivating students' CSE in college is vital for societal and individual development.

Studies prove that CSE is artificial due to several factors, including age, university background, and teacher guidance. For instance, a creative learning atmosphere has been found to improve students' academic performance, self-confidence, adaptability, problem-solving ability, interpersonal skills, and attendance percentage⁹. Although the above studies identified factors that manipulate CSE, the predictive roles of university members' mental perceptions and Academic Self-Efficacy (ASE) on CSE have not received sufficient scholarly attention¹¹. Moreover, extant studies have mainly focused on middle and higher-education students, with limited exploration of CSE among university students. The Decision Tree Model (DTM) is an effective data mining algorithm in Machine Learning (ML)¹². It analyses complex decision-making procedures as easier solutions to explain^13,14,15. Using a DTM, the researchers successfully predicted the elements that affect the success or failure of manufacturing innovation in China, as well as the response percentage to a survey on user satisfaction, attitude, and loyalty^16,17. This research aims to assess the effectiveness of using DTM in predicting CSE among college students and to identify the variables that predict CSE.

This study employs the Decision Tree Model (DTM) to predict key factors influencing Creative Self-Efficacy (CSE) among college students. DTM is selected for its interpretability, ability to handle mixed data types, and effectiveness in modeling non-linear relationships. Unlike black-box models, DTM offers transparent, rule-based outputs, enabling clear identification of predictor variables. Prior research in education and psychology supports its utility in analyzing complex decision-making scenarios. Its visual structure aids in practical interpretation, making it suitable for academic environments. DTM is ideal for researchers seeking an interpretable and data-driven approach to model behavioral or psychological outcomes with high clarity and relevance.

Creative Self-Efficacy
CSE is an individual's subjective decision about their capacity to complete creative activities successfully. It represents a domain-specific index of general self-efficacy and is an essential influencing factor and psychological basis for individuals' tasks and implementation of creative endeavors. Tierney and Farmer^18,19 introduced the Pygmalion CSE application model, asserting that CSE, an internal motivator of creative activity, effectively adopts individual creativity and significantly positively predicts an individual's capacity for creative skills. Furthermore, scholars highlight that CSE mirrors an individual's self-confidence in maintaining creative skills, as well as their abilities in learning and life processes²⁰. Studies indicate that people with a higher sense of CSE tend to focus on themselves more, opting for beliefs and goals that align with their interests and challenges to experience self-worth and achievement^21,22. While previous research has primarily emphasized CSE as an essential component of creative self-beliefs in realizing individual creative potential^22,23,24, this study defines college students' CSE as beliefs about their skill to complete specific inventive behaviors successfully.

Measurement and model of CSE
CSE is generally measured with scales adopted and adapted from previously developed SDQ devices. For example, it was measured by adapting Bandura's self-efficacy scale²⁵, Amabile's creativity scale²⁶, and Woodman, Sawyer, and Griffin's scale²⁷. Case studies, observations, and interviews have also been employed to discover CSE in the everyday behaviors of individuals within specific groups. However, a considerable portion of the study on CSE has centered on businesses and their employees. The application of these research studies' hypothetical ideas to other cohorts is limited. The study presented that the support of teachers and the method of a positive classroom environment are integral to increasing CSE among students²⁸.

A teacher's role involves imparting knowledge, a process inherently involving creativity. It concluded that increased CSE contributes to individuals' deeper involvement in creative skills, leading to better creative performance^29,30. Regarding age, significant gender-based differences in CSE were found among higher education students, while the CSE of adult subjects was significantly higher than that of older subjects. This proposes that the effect of age is not consistent. Creative self-concept theory informs the understanding of CSE, which encompasses the evaluation and awareness of one's performance in creative skills, categorized into CSE and creative self-identity^31,32,33,34. A creative self-concept has a significant impact on creative motivation, behavior, and outcomes. Based on this theory, researchers have developed structured models to study creative self-concept and creative skills in educational and work environments³⁵. Additionally, Amabile proposed a component model of creative behavior in 1983, identifying creative ability, motivation, and a creative environment as the three required conditions for creative behavior.

Predictive analysis methods for CSE
In the case of CSE, the primary predictive analysis methods are regression analysis, Confirmatory Factor Analysis (CFA), Principal Component Analysis (PCA), and meta-analysis. Using structural equation modeling and CFA, Tierney and Farmer found that creative skill, motivation, and environment predict individuals' creative skills and CSE. The work used PCA to demonstrate that CSE and creative self-identity constitute 2-D of creative self-concept, impacting creative motivation. Researchers who applied multiple regression modelling propose that creative skills, goals, emotions, and beliefs are vital predictors of CSE³⁶.

Predictors of CSE
Within existing academic works, several studies have focused on college students' CSE or utilized DTM to predict CSE^37,38. Therefore, this research aims to explore the application of DTM to predict college students' CSE. To achieve this, the study began by identifying predictors of CSE in the literature (Figure 1).

First, this research proposes perceptual behavior and environmental behavior as predictor factors of CSE. Planned behavior theory posits that subjective norms, attitudes, and perceived behavioral control influence behavioral intentions. Jointly determined by control opinions and perceived assistance factors, perceived behavioral control refers to the extent to which a person perceives that they are in control of the behavior they engage in; this implies that performing IB can enhance one's sense of control over creativity, which results from a sense of CSE³⁷. A study demonstrated that creativity is a predictor of CSE. CSE is the self-confidence that one has in one's ability to make creative skills, engage in creative behaviors, and achieve creative results. These studies also demonstrated a positive connection between the Psychological Sense of School Membership (PSSM) and CSE³⁸. Therefore, CSE may be adopted by various factors in the university environment to foster students' creativity³⁹. In terms of perceptual factors, this study proposes PR and ASE as predictors of CSE. Research has shown that PR is a person's ability to maintain stable emotional and behavioral responses when facing stress, setbacks, and adversity and to return to a normal state quickly. Accordingly, PR is significantly and positively related to emotional creativity, CSE, and general self-efficacy⁴⁰. Finally, it has been demonstrated that CSE can be predicted by ASE, which is influenced by a person's decision and beliefs about their ability to implement academic tasks^41,42. In summary, this study proposes four predictors of CSE: PR, ASE, PSSM, and IB.

Protocol

The research was tested at Chengdu normal university, SiChuan Province, China, with over 17,000 students. The SDQ resources and survey design were approved by the Ethics Review Committee of Chengdu Normal University, SiChuan Province, 611130, China. The university conducted its survey from September 10 to 13, 2023. Voluntary participation was ensured by obtaining informed consent

1. Enrollment of participants

1. Select a government university through a class-conscious random sampling method. Use the following criteria: (i) education for all follows the same plan and tests students in the same method; (ii) teaching is in the same language (Chinese); and (iii) each university has an optimal number of students to each teacher. Here, the Chengdu normal university, SiChuan province, China, was selected.
2. Select students aged 18 to 21 years as they understand their actions more clearly than ever before.
3. Before designing the SDQ tool, conduct investigative focus interviews with 5 students who agreed to participate in identifying probable predictive correlations between CSE and the recommended predictors. Subsequently, use convenience sampling to engage willing students.
4. Select all verified students. A total of 607 students verified all SDQs and confirmed their validity after thorough checking. The respondents' demographic data showed that the Male-to-Female ratio was 25.5% to 74.5%. Regarding residential background , 71.4% of the participants were from rural areas, while 28.6% came from urban settings. All students were pursuing liberal arts education under a standardized curriculum framework. The majority of participants belonged to low- to middle-income households, reflecting the broader socio-economic composition of public university enrollees in the region.
   NOTE: Typically, students entering teacher training higher education in China lean towards the liberal arts in their college entrance exams. Since female students generally display a greater interest in creative activities than their male counterparts, there is generally a higher percentage of female students in these universities, which explains the significant predominance of females in the sample.
5. Group participants by gender and grade to ensure the model was equitable. This enhances the results' applicability to most government-operated universities in Chengdu, which have similar socio-demographic populations.

2. Instrumentation

NOTE: This study used an SDQ comprising 6 sections containing 52 items, covering demographic information, the CSE Scale, the Psychological Resilience Scale, the Psychological Sense of School Membership Scale, the Academic Self-Efficacy Scale, and the Creative Behavior Scale. Demographic data includes gender, age, and whether the individual is from an urban or rural background. The English versions of the scales are translated into Chinese.

1. Apply the reverse translation method⁴³. Implement corrections and optimizations before finalizing the SDQ and replace the original scoring scale with a 5-point scale to maintain reliability while ensuring heterogeneity with other scales, thereby ensuring scale equivalence.
2. Measure all hypothesis items on a 5-point scale from 1=Strongly Disagree to 5=Strongly Agree.

3. Academic self-efficacy

1. Use the College Academic Self-Efficacy Scale (CASES) that was⁴⁴ specifically designed to assess self-efficacy among Chengdu normal university's students in academic training. The CASES comprises nine objects (e.g., I would like to do better than the other students; I am sure I can understand what the teacher is training; I want to do well in class). The reliability coefficient for this scale is 0.897.

4. Innovative behavior scale

1. To measure IB, use Innovative Behavior Scale⁴⁵. Maintain the scale's 3-D as follows: age group of innovative skills (e.g., When I am in trouble, I frequently get innovative skills), promotion of innovative skills (e.g., I will seek managerial support for innovative skills), and application of innovative skills (e.g., I will turn the innovative skill into a possible training). The reliability coefficient for this scale is 0.917.

5. Psychological sense of school membership scale

1. Use the PSSM to evaluate students' sense of belonging to a university regarding their emotional state, reactions, proof of identity, attachment, and behavioral attitude.
2. Developed by Goodenow⁴⁶, it contains 18 projects and has been generally translated into various languages, including Spanish and Chinese. Use the Chinese version of PSSM with 18 items (e.g., I feel like I belong to this university; Students take my opinions seriously)⁴⁷. The reliability coefficient for this scale is 0.853.

6. Psychological resilience scale

1. Use the 10 objects of the Connor-Davidson Resilience Scale (CD-RISC-10) to measure PR⁴⁸. Initially, with 25 objects, Campbell-Sills and Stein⁴⁹ improved the PR Scale to 10 objects overall in a shorter version.
2. Use the scale to measure an individual's feelings, reactions, and friendliness about their skill to heal from negative skills and flexibly adapt to variable external factors. It comprises of 5D, namely, the ability to relate (e.g., I can be flexible and adapt when change occurs), tolerance of negative affect (e.g., I can cope with problems with humor), acceptance of the change (e.g., I am stronger through skill), control (e.g., I can deal with upset emotions like anger), and mental impact (e.g., I can focus on my thought process when under pressure)⁴⁹. The reliability coefficient for this scale is 0.938.

7. CSE scale

1. Use the CSE scale to measure CSE, evaluate how much a person believes they can create creative skills⁵⁰. Condense the scale into three objects for the survey of Chinese students. The reliability coefficient for this scale is 0.860.

8. Data collection

1. This SQD research aimed to reveal the predictive factors of CSE among university students. Use an SDQ to measure college students' level of CSE along with four predictors⁵¹.
2. During this period, the head teacher encouraged students to voluntarily complete the SDQ to display the QR code at the end of each university day. Tell the participants that they can access the SDQ by scanning the QR code, answering the questions, and then clicking Submit.
3. Importantly, before students scan the QR code, instruct the head teacher to explain the study's aim to students and parents. Ensure voluntary participation by finding informed consent.
4. Administer the SDQ to the targeted student population. Ensure each participant responds to all mandatory objects. Validate the completeness and consistency of responses.
5. Define core constructs such as ASE, Parental Resilience (PR), and CSE. Code responses using a 5-point Likert scale ranging from 1 to 5. Assign participants to high or low types using a 60% or more threshold of the maximum scale score.
6. Divide the dataset into training and testing subsets. Use 70% of the data (n=423) for training the model. Allocate the remaining 30% (n=184) for testing purposes.
7. Apply the Classification and Regression Tree (CART) to find significant predictors. Use Gini impurity to determine node splits. Continue splitting until reaching a minimum number of cases or zero impurity.
8. Evaluate nodes for pruning using Mean Square Error (MSE). Truncate a node if pruning reduces the MSE. Retain a node if MSE increases after trimming.
9. Use the testing dataset to validate the DTM. Compute performance metrics such as accuracy, sensitivity, and specificity. Analyze misclassified cases to find DTM weaknesses.
10. Perform Pearson product-moment correlation analysis to study inter-variable relationships. Interpret the strength and direction of correlations based on correlation coefficient values.
11. Use the decision tree rules to identify how PR and ASE predict CSE levels. Quantify the contribution of each variable to classification outcomes.
12. Expand the feature set if the model accuracy falls below the desired threshold. Retrain the model with additional variables or a larger sample size to improve its accuracy.
13. Test the model's generalizability using data from different regions or institutions. Replicate the study protocol to validate findings across independent samples.
14. Each SDQ administration took ~25 min. Use an environmental control to ensure ambient noise <45 dB and temperature at 26-28 °C.
15. Train one facilitator for every 15 students, through a 2 h SDQ briefing session⁵². With the help of standard operating scripts, the facilitators were able to reduce the impact of their own opinions on the results⁵³. Ask the helper to read all the instructions exactly as written, avoid clarifying more than the necessary steps, and not attempt to explain the test questions. The facilitators helped eliminate the impact of interviewers and ensured that data came only from students' evaluations⁵⁴.
    NOTE: These additions address variability in implementation and enhance cross-contextual reproducibility. Since most students did not have personal devices and only a fraction of the universities had working Wi-Fi⁵⁵, administration on paper was the best solution. Moreover, research studies have shown no notable differences in psychometric results between SDQ digital and paper formats among individuals with lower literacy levels⁵⁶. It would be helpful in upcoming studies to examine how digitalization operates in private universities with strong resources⁵⁷. As no coding was done in this opinion, inter-rater reliability could not be determined.

9. Data analysis

1. Use a data mining method to analyze the survey data⁵⁸. Data mining can reveal patterns of student CSE and find three types of learners: integrative, goal-oriented, and sampling. This study opted for DTM as it has been widely used to predict students' academic performance and learning behaviors and can generate easily understandable rules⁵⁹.
   NOTE: Specifically, a Decision Tree (DT) builds classification rules from training samples and produces an easy-to-understand top-down graph consisting of a root node, several internal nodes, and multiple leaf nodes. Both root and internal nodes represent the corresponding test conditions (i.e., classification measures)⁶⁰, while the leaf nodes represent the final output. Rules can be inferred based on the tree model formed by each node. Moreover, DT proves a good tolerance for multicollinearity and can handle complex correlations between predictors. Additionally, definite DT is employed when predictor variables are categorical, while regression DT is suitable for continuous predictor variables⁶⁰.
2. In order to determine whether students' CSE is high or low, create a predictive model using a definite DT and analyze the significance of each factor in predicting CSE.

10. DT construction

1. Determine the splitting thresholds and optimal branching variables for the DT based on the rate of decrease of information entropy, which indicates the extent of impurity of the data set and is defined as Equation (1).
       (1)
   where D is the training dataset's sample size, m, P_k is the odds ratio for each type.
2. The information gain ratio measures the variance in information entropy of the dataset under many classification methods. Compute the gain ratio for the dataset D, partitioned into n subsets by the selected variable C, defined by Equation (2).
       (2)
3. In C5.0, select the attribute with the most significant information gain ratio as the split point, and form the multiple branches based on the value of that attribute to attain multiple subsets. Iterate the separation process until the last subset contains just the same type of data, thus realizing the inductive labelling of data.

11. DT pruning

1. Using the pruning method, prune leaf nodes layer by layer. After the DT is built, recurse the data sets to each leaf node according to the trained DTM.
2. Compute the MSE for leafy and leafless datasets. Truncate nodes if the MSE decreases after trimming; otherwise, preserve the nodes.

12. DT evaluation

1. Utilize 70% of the sample data (n=423) for training and the remaining 30% (n=184) for testing. The indicators for evaluating the quality of a model are accuracy, precision, and recall. Accuracy is defined as the percentage of correctly categorized samples out of all samples; precision is how many positive samples are real positive samples in the prediction result, and recall is how many positive samples are correctly predicted in the real samples.

13. Data analysis

1. Analyze the study results using a statistical analysis software. First, analyze CSE's frequency and concentration trends, as well as predictor factors, using descriptive statistical analyses. Second, perform DT analysis using the C5.0 algorithm, which improves upon ID3 and C4.5 as it runs faster and generates more robust predictive effects.

14. Data encoding

1. Using 60% as a differentiation point, label the sample into two groups: high CSE and low CSE. For the predictors, PR, IB, ASE, and PSSM were converted into binary variables according to specific principles (Table 1).

Results

Descriptive statistics
This study analyzed descriptive statistics for continuous and sequential variables (Table 2). Each variable is coded based on the principle that values exceeding 60% of the full scale indicate an above average performance. Notably, college students appear to have above-average CSE, with a mean value of 3.66 and a standard deviation of 0.80, which exceeds 60% of the full scale. This proposes that most college students have high CSE. The mean values of the four p...

Discussion

Using DT-C5.0, this study constructed a four-factor predictive model of college students' CSE to explore the role of these factors.All conclusions drawn in the study were based on the literature and relevant data analyses.

The effectiveness of the DTM in predicting levels of CSE was verified in the evaluation results. The model's accuracy, precision, and recall exceeded the 75 % criterion. The model has an Area Under the Curve (AUC) value of 0.762, surpassing 0.5, indicating i...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
2 TB HDD	Segante	6G SATA 7.2k 3.5"(LFF)	Storage
8 GB RAM	Kingston Technology	4Rx4 DDR4 LRDIMM 2666 MT/s.	Storage
Intel Core i5	Intel	Intel Core i7	Processor
NVIDIA GeForce RTX 4060 Ti	Nvidia Corporation	NA	Graphics Card
R		(version 4.1.2)	Programming Language
SPSS	IBM	30	Data Visualization Tool