Machine Learning in Education: Predicting Student Performance and Guiding Institutional Decisions

Abstract

Using Machine Learning (ML) in educational management transforms higher education strategy. This study examines students’ views on machine learning (ML) technologies and how they might be used to plan, monitor, and predict student performance. The Faculty of Industrial Engineering and Robotics surveyed 118 third-year undergraduates. It featured closed- and open-ended questions to collect quantitative and qualitative data. Descriptive statistics showed broad patterns, inferential tests (Chi-square, t-test, ANOVA) showed group differences, regression models predicted school outcomes, and exploratory factor analysis (EFA) and clustering found hidden attitudes and student profiles. A multi-method quantitative approach combining descriptive statistics, inferential tests, regression modeling, and exploratory techniques (EFA and clustering) was employed. The findings show that most students realize that ML may help them be more productive, adapt their study pathways, and learn about the future. Concerns remain regarding its accuracy, overreliance, and morality. The findings indicate that ML can both support and challenge educational management, depending on how responsibly it is implemented. Results show that institutions may utilize ML as a strategic tool to boost academic progress and make better judgments, provided they incorporate it responsibly and follow ethical rules and training.

1. Introduction

Opportunities to make data-based choices, tailor learning, and strategically administer this in institutions have opened up thanks to machine learning (ML) in higher education. Machine learning (ML) is more practical than artificial intelligence (AI) in schools due to its ability to evaluate and forecast data. According to Namoun and Alshanqiti (2020), this helps institutions understand their students’ needs, identify at-risk pupils, and streamline administrative work.

ML algorithms aid in evidence-based interventions, personalized learning, and efficient use of educational resources by uncovering hidden patterns in student data. Modern school administration is rapidly using ML due to the new data-driven insights it provides to educators (Ersozlu et al., 2024). The advantages of ML for adaptive learning, risk prediction, and improving institutions have been supported by recent research (Huang et al., 2025; Rodríguez-Ortiz et al., 2025).

Considering these findings, more recent studies have investigated the potential of AI and ML to enhance educational assessment, automate feedback systems, and encourage student engagement (Hwang et al., 2020; Holmes & Tuomi, 2022). The use of ML to school administration is yet underexplored, particularly for predictive analytics and institutional decision-making. Most of the research performed so far (Vieriu & Petrea, 2025) has focused on how AI can help teachers. Not many have looked at how data-driven systems can help with bigger educational goals, such as how to use resources, create student support programs, or make rules for schools. There has not been enough empirical research on how machine learning can be used as a management tool to predict student achievement and improve the efficiency of institutions.

Along with this growing interest, machine learning-enabled learning analytics are being used more to help academic management make decisions by using dashboards, guidance systems, and prescriptive analytics. Dashboards help teachers and administrators see changes in student performance right away, and prescriptive systems go even further by suggesting specific interventions and the best ways to use resources. These methods show how insights based on data can be directly used by managers, bridging the gap between learning analytics and institutional governance (Masiello et al., 2024; Tzimas & Demetriadis, 2024; Manly, 2024).

More studies show that clickstream data and LMS interactions can accurately predict how well students will do in school, how long they will stick with it, and how engaged they will be. Finding at-risk students and making retention plans based on data is made possible by these predictive models (Liu et al., 2022; Arévalo-Cordovilla & Peña, 2024) that are based on real-world data. Such uses show that ML is useful for managers in more ways than just teaching, making it a strategic tool for planning ahead in institutions.

But for institutions to use predictive systems, they need to be shown to be fair, not affected by how the data is handled, and easy for stakeholders to understand. These days, new studies show that imputation strategies can change performance metrics for subgroups and that post hoc explainability methods (like SHAP) can reveal policy-relevant outcome drivers. Additional research on clear admissions and performance models shows steps and proof that match predictive analytics with academic values and goals for fairness (Raftopoulos et al., 2024).

These issues highlight the significance of understanding public sentiment and confidence in ML-driven systems prior to their widespread adoption.

Along with predictive modeling, clustering methods help schools make decisions by revealing unique student groups that react differently to different types of academic support. By dividing students into groups based on how engaged they are, how they use resources, or how hard they think the problems are, managers can tailor their help and more accurately predict the results of the whole program. Clustering makes predictions more accurate and gives us a better picture of how different students are when used with multi-model fusion techniques (Pansri et al., 2024; Zou et al., 2025; Tudor et al., 2025).

ML-driven educational management combines learning analytics, organizational strategy, and data science on a theoretical level. In practice, predictive models based on student interaction data show early signs of performance or disengagement, allowing teachers and administrators to step in at the right time (Tsai et al., 2018; Albreiki et al., 2021). Also, clustering and regression methods have been used in other educational settings to find unique student groups and the factors that predict how well they will do. Using these methods in higher education administration can improve evidence-based decision-making and help institutions become more flexible and are always getting better (Mohamed Nafuri et al., 2022; Alwarthan et al., 2022). Implementing these strategies in the administration of higher education may help institutions become more adaptable and capable of continuously improving their practices, as well as enhance evidence-based decision-making.

Even with these opportunities, there remain issues. According to Nezami et al. (2024) and Kesgin et al. (2025), there are ethical and educational concerns with algorithmic transparency, data privacy, and over-reliance on automated methods. Students’ perceptions of ML tools’ accuracy, usefulness, and fairness provide light on their usage and trust in the classroom. To create ML-based systems that are both effective and morally compatible with educational principles, institutions must address these mindsets.

The purpose of this research is to gauge the sentiments and opinions of undergraduates on the use of machine learning in the field of educational administration. Students’ perceptions of the utility, accuracy, and ethical implications of machine learning tools, as well as their frequency and motivation for using them, are the three foci of this research. The third area of inquiry is the relationship between these factors and students reported academic performance. Using inferential statistics, regression analysis, exploratory factor analysis, and clustering techniques, this study aims to build a full model of how students feel about machine learning (ML) and what that means for educational management.

There is already a lot of research on AI in education, but this study adds to it by looking at ML as a management and prediction tool for making decisions at an institution. It also uses a multi-method quantitative framework to show how student perceptions affect academic outcomes and gives real-life examples from a technical university.

Research Objectives and Questions

This study aims to examine undergraduate students’ perceptions of machine learning (ML) tools in higher education and to explore how these perceptions relate to self-reported academic performance and institutional decision-making contexts. The study addresses the following research questions:

RQ1: What are students’ perceptions of ML in terms of usefulness, ease of use, trust, ethical concerns, over-reliance, institutional support, and intention to use?

RQ2: Which perception constructs predict students’ self-reported improvement in academic performance?

RQ3: Are there distinct student profiles based on these perception constructs?

2. Materials and Methods

2.1. Participants and Context

The study involved third-year undergraduate students from the Faculty of Industrial Engineering and Robotics at the National University of Science and Technology POLITEHNICA Bucharest.

A total of 118 responses were initially collected through convenience sampling. Following data quality screening, four responses were excluded for failing the attention-check item (Item 10: “Choose ‘Often’ for this question”), resulting in 114 valid cases included in the analyses. This procedure ensured that only attentive and reliable responses were retained, consistent with recommendations for survey data quality control (Meade & Craig, 2012).

Third-year students were selected because they possess sufficient academic and technical experience to evaluate the educational use of ML tools, while still being actively engaged in core university learning processes.

This non-probabilistic sampling method was deemed suitable for finding representative attitudes and use patterns within the target cohort due to the exploratory nature of the research and the homogeneity of the population (same faculty and study year).

Everyone who took part did it voluntarily and anonymously. There was zero collection of personally identifiable information. By choosing to fill out the survey, participants implicitly provided their informed permission.

This cohort improved the internal validity of the study by investigating attitudes among a population that is heavily focused on technology, as well as by ensuring academic uniformity and sufficient technical competence for evaluating ML-based educational apps.

2.2. Survey Design and Procedure

An online questionnaire created in Google Forms and disseminated via institutional communication channels was used to gather data.

The questionnaire was tailored to meet the research aims of this study by drawing on previous works on artificial intelligence and machine learning in education (Liu et al., 2022; Nezami et al., 2024; Pansri et al., 2024; Vieriu & Petrea, 2025).

There was a total of 34 items broken down into the following categories:

• Academic background and study habits (Items 1–2): average grade and weekly study hours.
• Use of AI/ML tools (Items 3–10): types of tools, frequency of use, activities, and verification behavior (including one attention-check item).
• Perceptions toward ML in education (Items 11–30):
  • Perceived Usefulness (PU) (Items 11–13),
  • Perceived Ease of Use (PEOU) (14–16),
  • Trust and Accuracy (TRUST) (17–19),
  • Ethical and Privacy Concerns (ETHICS) (20–22),
  • Over-Reliance and Critical Thinking (OVERREL) (23–25),
  • Institutional Support (SUPPORT) (26–28),
  • Intention to Use (INTENT) (29–30).

All these items were measured on a five-point Likert scale (1 = Strongly disagree to 5 = Strongly agree).

4.

Self-reported impact of ML (Items 31–32): perceived changes in performance and collaboration.

Items 31 and 32 were measured on a five-point Likert scale (1 = strongly disagree to 5 = strongly agree). Item 31 captured students’ self-reported perceived improvement in academic performance since starting to use AI/ML tools.

5.

Open-ended questions (Items 33–34): examples of ML use in learning and recommendations for integration into educational management.

To make sure the instrument was clear and reliable, it was pre-tested with 10 pupils. The items were somewhat revised during pilot testing to make them clearer and to make sure they were in line with the intended constructions.

The psychometric qualities of all scales, which were modified from previously validated instruments, were good (α = 0.49–0.83; ω = 0.50–0.83, where applicable).

To make the items easier to read and to bring them in line with other approved instruments, we made some little language tweaks.

The blank version of the questionnaire and the anonymized dataset of student responses are provided as Supplementary Materials.

2.3. Analytical Strategy and Statistical Procedures

The study used a quantitative multi-method strategy that integrated descriptive and inferential statistics with exploratory methods.

Descriptive statistics (mean, median, standard deviation, frequency) were used to summarize responses for each item and factor.

Inferential analyses examined relationships and differences:

• Chi-square tests (χ²) for associations between categorical variables (e.g., ML usage frequency and perceived performance).
• Independent-sample t-tests and one-way ANOVA for differences in perception factors across study-hour groups and performance bands.

Exploratory Factor Analysis (EFA) was conducted on the 20 Likert-scale perception items (11–30) to identify latent constructs and verify the dimensionality of the proposed factors (PU, PEOU, TRUST, ETHICS, OVERREL, SUPPORT, INTENT). Sampling adequacy was assessed via the Kaiser–Meyer–Olkin (KMO) test (more than 0.70) and Bartlett’s test of sphericity (p < 0.05). Principal Axis Factoring with oblimin rotation was applied. Items with loadings ≥ 0.40 and cross-loadings < 0.30 were retained. Reliability was tested using Cronbach’s alpha and McDonald’s omega (threshold ≥ 0.70).

Regression analysis was performed to predict self-reported academic performance (Item 31) from the factor scores (PU, TRUST, PEOU, SUPPORT, etc.), controlling study hours and usage frequency.

For each of the seven perception constructs, composite scores were computed as the arithmetic mean of the corresponding Likert-scale items (1 = strongly disagree to 5 = strongly agree). Specifically, Perceived Usefulness (PU_mean) was calculated as the mean of Items 11–13, Perceived Ease of Use (PEOU_mean) as the mean of Items 14–16, Trust and Accuracy (TRUST_mean) as the mean of Items 17–19, Ethical Concerns (ETH-ICS_mean) as the mean of Items 20–22, Over-Reliance (OVERREL_mean) as the mean of Items 23–25, Institutional Support (SUPPORT_mean) as the mean of Items 26–28, and Intention to Use (INTENT_mean) as the mean of Items 29–30. No reverse-coded items were included. When missing responses occurred, construct scores were calculated using the mean of the available items for that construct, following a pairwise deletion approach. These composite variables were subsequently used in all inferential, regression, and clustering analyses.

K-means clustering identified groups of students with similar profiles of ML attitudes and behaviors, based on standardized factor scores and usage frequency (Jain, 2010). The optimal number of clusters was determined using the elbow and silhouette methods, consistent with current clustering validation practices (Aggarwal & Reddy, 2014). Between-cluster differences on outcomes and covariates were examined using ANOVA and χ² tests, following conventional statistical procedures for cluster validation (Love et al., 2019). All analyses were conducted in JASP (v.0.19.3). Statistical significance was set at p < 0.05.

To ensure data quality, responses with more than 20% missing and those failing the attention-check (Item 10) were excluded; this procedure resulted in 114 valid cases used for statistical analysis, following best practices for inattentive responding and missing-data screening in survey research (Meade & Craig, 2012). Following the six-step approach proposed by Braun and Clarke (2006), thematic analysis was used to add qualitative context to the quantitative results in the open-ended replies (Items 33–34) of the survey.

Students’ views of ML and its relationship to their academic behaviors and participation may be fully understood using this integrated analytical method that combines exploratory, inferential, and predictive approaches.

To ensure statistical power, we used pairwise deletion to handle missing data and double-checked all statistical assumptions before analysis.

3. Results

3.1. Reliability and Internal Consistency

Using Cronbach’s α and McDonald’s ω on the valid dataset (N = 114), the internal consistency of each construct was assessed.

As shown in

Table 1. Reliability coefficients (Cronbach’s α and McDonald’s ω) for the perception constructs (N = 114).

Construct	Items	Cronbach’s α	McDonald’s ω
Perceived Usefulness (PU)	11–13	0.83	0.83
Perceived Ease of Use (PEOU)	14–16	0.65	0.66
Trust and Accuracy (TRUST)	17–19	0.82	0.82
Ethical Concerns (ETHICS)	20–22	0.70	0.72
Over-Reliance (OVERREL)	23–25	0.49	-
Institutional Support (SUPPORT)	26–28	0.73	0.74
Intention to Use (INTENT)	29–30	0.51	-

Note: α = Cronbach’s alpha; ω = McDonald’s omega; Values ≥ 0.70 indicate acceptable internal consistency for research purposes.

, most scales demonstrated satisfactory to excellent reliability, with α values ranging from 0.65 (Perceived Ease of Use) to 0.83 (Perceived Usefulness, Trust and Accuracy).

Given the small number of items per scale, lower Cronbach’s alpha values may occur even when items are conceptually coherent. Therefore, results involving the Over-Reliance and Intention to Use constructs are interpreted cautiously, and future studies should consider refining or expanding these scales.

McDonald’s ω coefficients were consistent with α values, confirming scale reliability.

The Ethical Concerns (α = 0.70, ω = 0.72) and Institutional Support (α = 0.73, ω = 0.74) constructs also showed acceptable reliability.

Lower internal consistency was observed for the Over-Reliance (α = 0.49) and Intention to Use (α = 0.51) scales, which may reflect heterogeneous item content and the short number of items per scale.

McDonald’s ω could not be estimated for the Over-Reliance and Intention to Use scales because of the limited number of items and moderate inter-item correlations. In such cases, Cronbach’s α provides a reliable estimate of internal consistency.

Overall, these results support the internal reliability of the measurement instrument.

3.2. Descriptive Results

Table 2. Descriptive statistics for the perception constructs (N = 114).

Construct	M	SD	Min	Max
Perceived Usefulness (PU)	3.61	0.89	1.33	5.00
Perceived Ease of Use (PEOU)	3.64	0.83	1.67	5.00
Trust and Accuracy (TRUST)	3.32	0.83	1.00	5.00
Ethical Concerns (ETHICS)	3.13	1.02	1.00	5.00
Over-Reliance (OVERREL)	3.47	0.85	1.33	5.00
Institutional Support (SUPPORT)	2.31	0.95	1.00	5.00
Intention to Use (INTENT)	4.19	0.81	2.00	5.00

Note: M = Mean; SD = Standard Deviation; Min = Minimum; Max = Maximum.

displays descriptive data for each of the seven perceptual components.

Perceived Ease of Use (M = 3.64, SD = 0.83) and Perceived Usefulness (M = 3.61, SD = 0.89) were reasonably high among students, indicating that they had good opinions regarding the accessibility and application of ML technologies in education.

The intermediate means for Trust and Accuracy (M = 3.32, SD = 0.83) and Ethical Concerns (M = 3.13, SD = 1.02) suggest that students are aware of the advantages of ML, but they are still wary of privacy and fairness problems.

The lowest mean was found for Institutional Support (M = 2.31, SD = 0.95), highlighting limited perceived guidance and resources provided by the university.

Conversely, Intention to Use recorded the highest score (M = 4.19, SD = 0.81), showing strong motivation among students to continue using ML responsibly.

All variables were measured on a five-point Likert scale (1 = Strongly disagree to 5 = Strongly agree).

3.3. Correlational Analysis

Pearson correlation coefficients among the seven perception constructs are presented in

Table 3. Pearson correlations among the perception constructs (N = 114).

Variable	PU	PEOU	TRUST	ETHICS	OVERREL	SUPPORT	INTENT
PU	-	0.51 ***	0.68 ***	−0.04	−0.09	0.10	0.55 ***
PEOU		-	0.61 ***	−0.05	0.09	0.08	0.56 ***
TRUST			-	−0.16	0.07	0.20 *	0.62 ***
ETHICS				-	0.36 ***	0.00	−0.17
OVERREL					-	0.03	0.11
SUPPORT						-	0.03
INTENT							-

Note: r = Pearson correlation coefficient; p < 0.05 *, p < 0.001 ***.

.

Strong positive correlations were found between Perceived Usefulness and Trust and Accuracy (r = 0.68, p < 0.001), as well as between Trust and Accuracy and Intention to Use (r = 0.62, p < 0.001).

Perceived Ease of Use was positively related to both Perceived Usefulness (r = 0.51, p < 0.001) and Trust and Accuracy (r = 0.61, p < 0.001), indicating that ease of use is closely associated with students’ confidence in ML tools.

There was a slight positive correlation between Trust and Institutional Support (r = 0.20, p < 0.05), and a substantial positive correlation between Over-Reliance and Ethical Concerns (r = 0.36, p < 0.001).

In general, our findings indicate that the desire to utilize ML technologies in academic settings is positively correlated with perceived utility, simplicity of use, and trust.

3.4. Regression Analysis

The elements that students perceived as influencing their self-reported improvement in academic performance (Item 31) were determined by a multiple linear regression analysis (see

Table 4. Multiple linear regression predicting self-reported academic performance (N = 114).

Predictor	β	t	p	VIF
Perceived Usefulness (PU)	0.23	2.10	0.039	2.15
Perceived Ease of Use (PEOU)	0.09	0.71	0.479	1.77
Trust and Accuracy (TRUST)	0.29	2.41	0.018	2.63
Ethical Concerns (ETHICS)	−0.05	−0.54	0.591	1.29
Over-Reliance (OVERREL)	−0.10	−1.17	0.247	1.30
Institutional Support (SUPPORT)	−0.09	−1.14	0.256	1.05
Intention to Use (INTENT)	0.17	1.65	0.101	1.96

Note: β = Standardized regression coefficient; p = significance level; VIF = Variance Inflation Factor.

).

The dependent variable used in the regression model (Item 31: perceived improvement in academic performance) ranged from 1 to 5, with higher values indicating greater perceived improvement.

It is important to note that the regression analysis models students’ self-reported perceived improvement rather than objectively measured academic grades or ML-based performance predictions.

The overall model was statistically significant (F(7,106) = 11.18, p < 0.001) explaining 42.5% of the variance in perceived academic performance (R² = 0.43, Adjusted R² = 0.39).

Students who see ML tools as both helpful and trustworthy are more likely to report better academic achievements, as shown by the significant positive indicators Perceived Usefulness (β = 0.23, p = 0.039) and Trust and Accuracy (β = 0.29, p = 0.018).

Notable absences from the statistical analysis include the following constructs: perceived ease of use, ethical concerns, over-reliance, institutional support, and intention to use.

The stability of the model was confirmed using multicollinearity diagnostics, which showed acceptable values (all VIF < 2.7).

3.5. Group Differences (ANOVA)

We ran seven one-way ANOVA tests, one for each perception construct (see

Table 5. One-way ANOVA results for study-hour differences in perception constructs (N = 114).

Construct	F(3,110)	p	${\mathit{\eta }}^2$	Post Hoc Tukey Results
Perceived Usefulness (PU)	3.57	0.017	0.09	10–14 h > 0–2 h; 10–14 h > 3–5 h
Perceived Ease of Use (PEOU)	1.86	0.140	0.05	-
Trust and Accuracy (TRUST)	2.53	0.061	0.06	-
Ethical Concerns (ETHICS)	0.95	0.419	0.03	-
Over-Reliance (OVERREL)	0.94	0.426	0.03	-
Institutional Support (SUPPORT)	2.85	0.041	0.07	3–9 h > 0–2 h (marginal)
Intention to Use (INTENT)	1.68	0.176	0.04	-

Note: Levene’s tests confirmed equal variances across all groups (p > 0.10). Significant results (p < 0.05) are present in Table 5. Post hoc Tukey tests revealed higher perceived usefulness and institutional support among students who studied 10–14 h per week compared to those studying less than 5 h.

), to explore whether students’ views of ML tools varied with their weekly study hours outside of class.

When looking at perceived usefulness and institutional support, substantial impacts were found.

The impact size of study hours on Perceived Usefulness (PU) was minor to moderate, as shown by a statistically significant effect (F(3,110) = 3.57, p = 0.017, η² = 0.09). Compared to students who studied for 0–2 h (p = 0.016) or 3–5 h (p = 0.040) per week, those who studied for 10–14 h per week reported significantly greater perceived utility, according to post hoc Tukey comparisons.

Likewise, there was a significant difference in Institutional Support (SUPPORT) ratings across study-hour groups (F(3,110) = 2.85, p = 0.041, η² = 0.07), with students studying 3–9 h per week achieving somewhat higher scores than those studying fewer.

Students who put in more time studying seemed to have greater faith in ML outputs when it came to Trust and Accuracy (F(3,110) = 2.53, p = 0.061), while the differences were not statistically significant (p > 0.05).

No significant differences were observed for Perceived Ease of Use (p = 0.140), Ethical Concerns (p = 0.419), Over-Reliance (p = 0.426), or Intention to Use (p = 0.176).

Levene’s tests confirmed homogeneity of variances for all variables (all p > 0.10).

Complementary t-tests and Chi-square analyses examining ML usage frequency did not reveal any significant group differences (p > 0.05).

3.6. Exploratory Factor Analysis (EFA)

An exploratory factor analysis (EFA) was conducted on the 20 perception items (Items 11–30) to identify the underlying dimensional structure of students’ perceptions of machine learning (ML) in education (see

Table 6. Exploratory factor analysis results for perception constructs (Items 11–30).

Factor	Eigenvalue	Variance (%)	Representative Items	Interpretation
Perceived Usefulness and Trust	4.79	23.9	11–19	Cognitive and instrumental benefits of ML tools
Institutional Support	1.60	8.0	26–28	Institutional and pedagogical backing
Ethical and Over-Reliance Concerns	1.58	7.9	20–25	Awareness of ethical and dependency risks
Ease of Interaction	1.23	6.2	14–16	Ease of learning and integrating ML
Intention to Use and Advocacy	1.15	5.8	29–30	Future behavioral intentions and recommendations
Total	-	51.8	-	-

Note: Extraction method: Minimum residual; Rotation: Oblimin; KMO = 0.79; Bartlett’s χ²(190) = 901.02, p < 0.001; Total variance explained = 51.8%.

).

The minimum residual extraction method with oblimin rotation was applied, using the correlation matrix as the basis of analysis. Sampling adequacy was confirmed by a satisfactory Kaiser–Meyer–Olkin (KMO) value of 0.79, and Bartlett’s test of sphericity was significant (χ²(190) = 901.02, p < 0.001), confirming that the data were suitable for factor analysis.

With eigenvalues larger than 1, five components were identified, which accounted for 51.8% of the overall variation.

A distinct and understandable factor structure, matching the conceptual characteristics of the investigation, was revealed by the rotating solution:

• Perceived Usefulness and Trust (Items 11–19)—reflecting the perceived benefits and reliability of ML in supporting academic activities.
• Institutional Support (Items 26–28)—describing the institutional and pedagogical support provided for responsible ML use.
• Ethical and Over-Reliance Concerns (Items 20–25)—capturing students’ awareness of privacy, bias, and dependency issues.
• Ease of Interaction (Items 14–16)—addressing students’ experience in using ML tools.
• Intention to Use and Advocacy (Items 29–30)—reflecting behavioral intention and recommendations to others.

All items loaded above 0.40 on their respective factors, with minimal cross-loadings, confirming the construct validity and internal coherence of the instrument.

Although seven theoretical constructs were initially proposed, the EFA indicated a more parsimonious five-factor solution, with Perceived Usefulness and Trust merging into a unified component, and Ethical and Over-Reliance items loading on a shared “risk awareness” factor.

3.7. Cluster Analysis

To identify distinct profiles of students based on their perceptions of machine learning (ML) tools, a hierarchical cluster analysis was conducted using Ward’s method and Euclidean distance on the seven perception constructs (Perceived Usefulness, Ease of Use, Trust, Ethical Concerns, Over-Reliance, Institutional Support, and Intention to Use) as can be shown in

Table 7. Cluster centers for student profiles based on ML perception constructs (standardized scores).

Factor	Cluster 1—Balanced Supporters	Cluster 2—Critical Evaluators	Cluster 3—Optimistic Users
Perceived Usefulness (PU)	0.06	0.40	−0.51
Perceived Ease of Use (PEOU)	0.86	−0.86	0.57
Trust and Accuracy (TRUST)	0.78	0.00	−0.41
Ethical Concerns (ETHICS)	0.65	−0.64	0.42
Over-Reliance (OVERREL)	0.56	−0.70	0.54
Institutional Support (SUPPORT)	−0.93	−0.12	0.63
Intention to Use (INTENT)	0.55	−0.70	0.55
Cluster size (n)	22	50	42

Note: Hierarchical clustering was conducted using Ward’s method and Euclidean distance on standardized variables. The three-cluster solution provided an interpretable segmentation of students’ ML perceptions (R² = 0.34, Silhouette = 0.16).

.

All variables were standardized prior to clustering to ensure comparability across dimensions.

The model fit indices supported a three-cluster solution (R² = 0.34, AIC = 565.59, BIC = 623.05, Silhouette = 0.16), providing the most interpretable and balanced grouping (see Table 7).

Cluster sizes were well distributed, with 22 participants in Cluster 1, 50 in Cluster 2, and 42 in Cluster 3.

Each cluster reflected a distinct perceptual profile of ML use in educational contexts:

• Cluster 1—Balanced Supporters (n = 22): Students reporting moderately high scores for Ease of Use, Trust, and Ethical Awareness (PEOU_mean = 0.86; TRUST_mean = 0.78; ETHICS_mean = 0.65), coupled with moderate Perceived Usefulness and Intention to Use (PU_mean = 0.06; INTENT_mean = 0.55).
• Cluster 2—Critical Evaluators (n = 50): The largest cluster, characterized by relatively low scores across most constructs, especially Ease of Use (PEOU_mean = −0.86) and Intention to Use (INTENT_mean = −0.70), but moderate perceived usefulness (PU_mean = 0.40).
• Cluster 3—Optimistic Users (n = 42): Participants showing moderate perceived usefulness but higher institutional support (PU_mean = −0.51; SUPPORT_mean = 0.63) and positive attitudes toward ethical and reliance dimensions (ETHICS_mean = 0.42; OVERREL_mean = 0.54).

Cluster analysis showed that various types of ML users engaged with the technology in different ways. Optimistic users saw ML as useful and well-supported, whereas Critical Evaluators were dubious about its simplicity of use and institutional support. The findings show that students’ attitudes of ML are diverse. The varied perspectives of students on the use of ML in the classroom are shown by this segmentation.

4. Discussion

By providing empirical insights from a technical university environment, this study adds to the expanding corpus of literature on the function of ML in educational administration. The findings provide light on the positive and negative aspects of ML integration in universities. Students’ generally favorable impressions of ML tools’ utility and user-friendliness reflect the growing acceptance of da-ta-driven technology in the classroom. Full acceptance will need institutional leadership and openness, nevertheless, because of moderate trust and strong worries about ethical difficulties.

Prior research has shown that students see ML and AI-based systems as useful tools to aid in learning and improve performance (Ersozlu et al., 2024; Hwang et al., 2020), and the high levels of Perceived Usefulness (M = 3.61) and Ease of Use (M = 3.64) are consistent with this perception. According to these results, the perception of ML applications as helpful to academic production increases as their accessibility and intuitiveness increase. Students acknowledge the promise of ML, but they are still wary of the dependability and interpretability of the data, as shown by the modest Trust and Accuracy (M = 3.32). Both (Nezami et al., 2024) and (Kesgin et al., 2025), have highlighted the need of educational prediction models being fair, explainable, and accountable; this am-bivalence is in line with their concerns.

Additional regression findings showed that self-reported enhancements in academic performance were substantially predicted by Perceived Usefulness and Trust and Accuracy. Pupils have more faith in their own abilities to succeed academically is a direct result of their faith in the validity and trustworthiness of ML tools. Research that has connected the use of technology with educational achievements has also shown similar patterns of prediction (Albreiki et al., 2021; Huang et al., 2025). There is a clear disparity between personal drive and institutional preparedness to use ML, as neither institutional support nor ethical concerns were important predictors. Previous requests for institutions to establish transparent plans for ML integration, including regulations for fairness, privacy, and responsible data use, have been echoed by this (Raftopoulos et al., 2024; Rodríguez-Ortiz et al., 2025).

These findings should be interpreted in the context of self-reported outcomes and reflect attitudinal predictors rather than direct algorithmic prediction of academic results.

Using a more concise five-factor framework, the exploratory factor analysis united trust and perceived utility into one construct and grouped ethical and dependence concerns into a common “risk awareness” dimension. This indicates that students see ML as an integrated technical event, where the reliability and practical advantages are interrelated. This confluence of ideas has been found in previous research on the topic of technology adoption in schools (Holmes & Tuomi, 2022; Manly, 2024), demonstrating how confidence in digital spaces is always changing.

Students’ varied perspectives on the use of ML in the classroom are reflected in the three unique profiles revealed by the cluster analysis: Balanced Supporters, Critical Evaluators, and Optimistic Users. A pragmatic approach to ML integration, rather than blind acceptance, was proposed by Balanced Supporters, who showed moderate confidence across all parameters. Masiello et al. (2024) found that confidence in learning analytics tools is typically undermined by inadequate institutional framing. Similarly, critical evaluators, who made up over half of the sample, exhibited skepticism toward ease of use and institutional support. On the other hand, optimistic users linked ML to efficiency, fairness, and institutional endorsement, demonstrating how data-driven management may, with the right backing, encourage participation and the perception of justice. The need of tailored institutional training and communication tactics is emphasized by this segmentation, which verifies the diversity of user attitudes.

The findings highlight that institutional support is still the least strong aspect (M = 2.31) from a management standpoint. Even if students are eager to utilize ML technologies, their adoption may be hindered by a lack of formal supervision, instructional frameworks, and ethical guarantees. Tsai et al. (2018) and Tzimas and Demetriadis (2024) both agree that in order to link predictive analytics with real decision-making and learner support procedures, institutional infrastructures are crucial. Hence, colleges should do more than just install ML systems; they should also make sure faculty and students understand the systems’ purpose, constraints, and guiding principles for administration.

By combining inferential, predictive, and exploratory studies, the multifaceted utility of ML in education is shown. Policy alignment, technological transparency, and ethical compliance are systemic elements that decide sustainability, whereas individual views determine short-term acceptability. The results should be further investigated in future studies by including bigger datasets from other institutions and by examining the long-term impacts of ML on measurable academic performance. To further understand how opinions change with time and as an organization develops, a mixed-method approach integrating performance data, interviews, and behavioral logs might be useful.

To sum up, this research highlights the importance of machine learning in school administration, highlighting its dual position as an area that needs rigorous ethical stewardship and a driver for efficiency and customization. Building student trust and realizing the full predictive and managerial potential of ML in higher education requires responsible and open deployment, backed by institutional leadership.

From a practical perspective, the identified perception factors and student profiles provide actionable insights for educational institutions. Universities may use these findings to tailor training programs, develop transparent guidelines for ML use, and design differentiated support strategies aligned with student profiles (e.g., critical evaluators versus optimistic users). Such evidence-based approaches can support responsible ML integration while strengthening trust, ethical awareness, and informed decision-making at the institutional level.

Limitations

This study is based on a convenience sample from a single faculty and institution, which limits the generalizability of the findings. The results rely on self-reported perceptions and outcomes, and the cross-sectional design does not allow causal inferences. In addition, two constructs exhibited lower internal consistency, and cluster separation was modest. These limitations should be considered when interpreting the results. Future re-search should employ larger, multi-institutional samples, longitudinal designs, and objective performance indicators.

5. Conclusions

This paper offers concrete examples of how ML might enhance strategic decision-making and educational management in the higher education sector. Students’ trust-based and ethical perceptions of ML shape their evaluations of its academic and institutional value, according to the study, which integrated descriptive, inferential, and exploratory analyses to show this.

It is clear from the results that trust in data dependability and openness as well as perceived utility are the most important factors in determining whether students report an increase in their academic performance. This highlights the fact that functional advantages are not enough on their own. There is a consistent disparity between individual acceptance and organizational preparation for ML integration; institutional support, on the other hand, appeared as the weakest element. The importance of readily available technology, governance frameworks, training, and clear institutional communication in ensuring the ethical and successful use of ML is highlighted by these observations.

Looking at ML from a management lens, the paper emphasizes how it can be both a boon to efficiency and customization and a bane to education and ethics. Therefore, in order to make sure that predictive algorithms are in line with educational principles and build confidence among stakeholders, universities should follow a responsible innovation approach, grounding ML-based decision-making in openness, equity, and inclusivity.

These results could inform future studies that use longitudinal designs and cross-institutional samples to examine the impact of exposure and policy changes on people’s perceptions of ML. One way to better understand the long-term effects of ML-driven tools on learning habits and organizational cultures is to combine quantitative analytics with qualitative research.

Ultimately, the research contributes to our knowledge of ML’s function in higher education by establishing a connection between student views and management realities. Responsible and well-communicated ML integration can serve as a cornerstone for data-informed, equitable, and adaptive educational management.

These conclusions should be interpreted considering the study’s exploratory design, reliance on self-reported measures, and the limited scope of the sample. While the findings offer useful insights into student perceptions of ML in higher education, they do not constitute direct prediction of academic outcomes or institutional performance.