Predicting Masteristic Learning with Machine Learning

Abstract

This review follows Lee's (2026) study of machine learning models that predict masteristic learning among 1,555 organizational employees. RandomForest performed best. Excellence in work, pride in work, and peer learning support were the strongest predictors. The discussion considers nonlinear relationships, how to interpret SHAP importance, and the distinction between prediction and causal explanation.

Introduction: Reversing the direction of prediction

Masterity theory describes contemporary masters who develop proficiency in their field of work, reach the highest levels of achievement, and share what they learn with those around them and their communities (Chang, 2015). Scales developed for four domains—masteristic learning, masteristic achievement, masteristic work, and masteristic influence—made it possible to test the theory empirically.

Empirical studies conducted since the scales were developed have mainly treated masteristic learning as a mediator or masteristic achievement as an outcome. Few studies have treated masteristic learning itself as the outcome to be predicted and examined how other components of masterity, together with organizational and individual contextual factors, explain it. A previous study (Lee, 2024) confirmed that masteristic learning predicts masteristic achievement. The study reverses the direction of prediction to examine the cyclical relationship between learning and achievement assumed by the masterity model.

The study addresses three research questions: which of the three ensemble methods performs best, how the predictors compare in relative importance, and what functional relationships exist between the predictors and masteristic learning.

Theoretical background: Machine learning and three levels of predictors

Drawing on Yarkoni and Westfall (2017), the study describes machine learning as a methodological alternative for research in the psychological and behavioral sciences. In these fields, regression analysis and structural equation modeling have emphasized causal explanation, with limited validation of out-of-sample predictive performance. Tree-based ensemble methods capture interactions and nonlinear relationships between variables without requiring prior assumptions about model structure.

The predictors are organized into three levels.

Level	Factors	Supporting arguments
Masterity factors	Masteristic achievement (excellence in work, work performance, pride in work)	The cyclical relationship between learning and achievement (Chang, 2015). Excellence provides a foundation for self-efficacy and elaborate self-reflection, performance triggers reflection in experiential learning through feedback, and pride strengthens exploration and goal pursuit through the broadening and building of positive emotions
Organizational factors	Organizational culture (clan, adhocracy, market, hierarchy), learning support environment (peer learning support, institutional learning support)	Informal learning studies have reported different directions of prediction across culture types. Previous findings on the learning support environment are available at the subfactor level
Individual factors	Gender, age, educational attainment, tenure, company size	—

The model predicts masteristic learning from three layers of factors; excellence in work, pride in work, and peer support for learning were the top predictors

The study by Lee (2023) found significant positive relationships between peer learning support and all five subfactors of masteristic learning, whereas institutional learning support was significantly related to only some subfactors. These findings led the author to expect informal, relationship-based support to predict masteristic learning more broadly and strongly than formal, institutional support. The two subfactors were entered separately.

The study also considers potential conceptual overlap between masteristic learning and masteristic achievement. The masterity model treats them as temporally ordered developmental stages. Their correlation (.700) is lower than the square root of the average variance extracted for each construct (.710, .728), meeting the criterion of Fornell and Larcker (1981).

Methods

Participants. The study population consists of wage employees aged 19 and older in South Korea. The data came from an online survey conducted in April–May 2021 by Professor Chang’s research team to develop a masterity measurement instrument. The initial sample included 2,531 respondents. Careless responses were removed using a multiple-hurdle approach. The first-stage agreement–disagreement consistency test excluded 786 respondents, and the second-stage straightlining check excluded 190, leaving 1,555 respondents for the final analysis.

Measures. All items use a 7-point scale.

Classification	Factor	Subfactors	Number of items	Source
Outcome factor	Masteristic learning	5	17	Jang et al. (2021)
Predictor	Masteristic achievement	3	20	Lee et al. (2021)
Predictor	Organizational culture	4	16	Cameron & Quinn (2011)
Predictor	Learning support environment	2	8	Na and Kim (2016)
Predictor	Individual factors	—	5	Gender, age, educational attainment, tenure, company size

Confirmatory factor analysis of the second-order factor structure yielded standardized loadings of .65–.93, construct reliability of .78–.89, and Cronbach’s α of .74–.89.

Predictive models and analytical procedures. The study compared a model with predictors entered at the component level (3 variables) with one using predictors at the subfactor level (14 variables). The subfactor-level model produced more accurate predictions and was selected as the final model. The data were split into training data (80%, n=1,244) and test data (20%, n=311). Hyperparameters were searched using 10-fold cross-validation and RandomizedSearchCV. The models were evaluated using RMSE, MAPE, and R². SHAP was used to assess predictor importance in the best-performing model, and partial dependence analysis examined whether functional relationships were linear or nonlinear.

Results

Predictive model performance. The standard deviation of masteristic learning was .792. RMSE values on the test data were lower than this for all three models, ranging from .556 to .570. MAPE ranged from 8.45 to 8.68%, placing all three models in the excellent category (10% or below). RandomForest performed best across all metrics—RMSE (.556), MAPE (8.45%), and R²—and was selected as the optimal model. The paper interprets an RMSE of .556 to mean that, when the actual masteristic learning score is 5, the predicted value falls approximately between 4.44 and 5.56.

Predictor importance (SHAP).

Rank	Variable	Category	SHAP value	Proportion
1	Excellence in work	Masteristic achievement	.215	29.64%
2	Pride in work	Masteristic achievement	.157	21.66%
3	Peer learning support	Learning support environment	.092	12.74%
4	Work performance	Masteristic achievement	.069	9.53%
5	Hierarchy culture	Organizational culture	.035	4.86%
6	Market culture	Organizational culture	.032	4.43%
7	Adhocracy culture	Organizational culture	.027	3.76%
8	Institutional learning support	Learning support environment	.025	3.42%
9	Clan culture	Organizational culture	.023	3.12%
10–14	Gender, company size, educational attainment, tenure, age	Individual factors	.008–.013	Less than 2% each

Two subdimensions of masteristic achievement and peer support account for 64% of the predictive importance for masteristic learning

The top four factors together accounted for 73.6% of predictive importance. The share for institutional learning support (3.42%) was approximately one-quarter of that for peer learning support (12.74%).

Relationships between predictors and masteristic learning. Partial dependence analysis showed that predicted masteristic learning increased linearly with excellence in work across the entire range. Peer learning support and pride in work showed nonlinear (stepwise) patterns: their predictive contributions rose sharply beyond approximately 4.5 points.

Conclusions and recommendations

The paper draws four conclusions.

First, masteristic achievement is the strongest predictor of masteristic learning. The paper treats this finding, together with previous research on the learning→achievement direction, as empirical support for a cyclical, mutually reinforcing relationship. Among the achievement subfactors, excellence in work and pride in work contributed more than work performance. The paper interprets masteristic learning as closely intertwined with proficiency acquired through personal effort and the pride that arises from it. Longitudinal data are necessary to establish the direction and strength of the relationship.

Second, peer learning support had greater predictive power than organizational culture. This supports the role of social and informal support emphasized by masterity theory. The paper recommends that organizations prioritize cultivating a culture of learning among peers.

Third, the four subfactors of organizational culture had similar levels of importance, in the 3–5% range. The paper suggests that the overall strength of perceptions may have played a greater role than culture type. Because the study did not measure this, it remains a hypothesis for future research.

Fourth, the paper interprets the nonlinear relationships as indicating that support and experience begin to facilitate learning once they accumulate to a certain level. Organizational interventions may therefore be more effective when focused on members below the threshold.

The paper calls for an integrated model incorporating masteristic work and masteristic influence, along with techniques such as causal forests. Other directions include testing cross-lagged effects with repeated-measures longitudinal data, using multilevel analysis to separate individual-level and organizational-level effects, and extending comparisons to other forms of work, including independent work.

Discussion

Using the same data source, the study by Choi et al. (2025) treated the learning support environment as a single construct and used structural equation modeling to test a pathway through job crafting to masteristic influence. Organizational culture and the learning support environment in the present study were measured through individual perceptions. The importance assigned to organizational culture therefore reflects how well those individual perceptions predict masteristic learning.

The study by Putka et al. (2018) compared modern prediction methods in organizational research with traditional regression-based methods. Its applied examples reported predictive performance alongside analyses of how the variables operate. The paper by Yarkoni and Westfall (2017) proposed giving out-of-sample prediction greater weight when evaluating psychological research. It drew on the principles of overfitting, cross-validation, and regularization. The present study puts that prediction-centered approach into practice by setting aside test data and conducting 10-fold cross-validation.

The paper by Shmueli (2010) argued that causal explanation and empirical prediction differ in purpose and evaluation criteria. This distinction extends from variable selection to model validation. The high importance of excellence in work and pride in work is a property of a model that predicts masteristic learning scores. The cyclical relationship between learning and achievement forms part of masterity theory’s developmental explanation. These data alone cannot test that cyclical relationship.

The study by Lundberg and Lee (2017) proposed SHAP, which assigns each feature a contribution to an individual prediction, and brought additive feature attribution methods into a common framework. SHAP importance indicates how much each variable contributes to the values predicted by a trained model. The gap between peer learning support (12.74%) and institutional learning support (3.42%) reflects a difference in their contributions within the model. The study by Janzing et al. (2020) examined the probability distributions assigned to features omitted from Shapley value calculations. It distinguished observational conditional probabilities from interventional probabilities. With correlated concepts, the two methods assign different levels of importance to the same variable.

The paper by Kapoor and Narayanan (2023) described how data splitting and preprocessing affect reported predictive performance.

References

이호진. (2026). 머신러닝 기법을 활용한 장인적 학습의 예측모형과 영향요인분석: 조직근로자를 중심으로. 인문사회과학연구, 8(4), 1129–1148. https://doi.org/10.62783/SHSS.8.4.65
Lee, H. (2026). Predictive model and influence factors analysis of masteristic learning using machine learning technique: Focusing on organizational employees. Studies on Humanities and Social Sciences, 8(4), 1129–1148. https://doi.org/10.62783/SHSS.8.4.65
Janzing, D., Minorics, L., & Blöbaum, P. (2020). Feature relevance quantification in explainable AI: A causal problem. In S. Chiappa & R. Calandra (Eds.), Proceedings of the twenty third international conference on artificial intelligence and statistics (Vol. 108, pp. 2907–2916). PMLR.
Kapoor, S., & Narayanan, A. (2023). Leakage and the reproducibility crisis in machine-learning-based science. Patterns, 4(9), Article 100804. https://doi.org/10.1016/j.patter.2023.100804
Lundberg, S. M., & Lee, S.-I. (2017). A unified approach to interpreting model predictions. In I. Guyon, U. von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, & R. Garnett (Eds.), Advances in neural information processing systems (Vol. 30, pp. 4765–4774). Curran Associates.
Putka, D. J., Beatty, A. S., & Reeder, M. C. (2018). Modern prediction methods: New perspectives on a common problem. Organizational Research Methods, 21(3), 689–732. https://doi.org/10.1177/1094428117697041
Shmueli, G. (2010). To explain or to predict? Statistical Science, 25(3), 289–310. https://doi.org/10.1214/10-STS330
Yarkoni, T., & Westfall, J. (2017). Choosing prediction over explanation in psychology: Lessons from machine learning. Perspectives on Psychological Science, 12(6), 1100–1122. https://doi.org/10.1177/1745691617693393