DEVELOPMENT OF AN AI-BASED PERSONALIZED LEARNING SYSTEM FOR PRESCHOOL EDUCATION WITH PARENTAL CONTENT CUSTOMIZATION AND INTERACTIVE ENGAGEMENT

УДК 004.8

ABSTRACT

Early childhood education represents a critical developmental period during which cognitive, linguistic, and socio-emotional foundations are established. However, traditional preschool instructional approaches often fail to address differences in developmental readiness, learning preferences, engagement patterns, and cultural backgrounds among children aged three to six. This study aims to develop and evaluate an AI-based Personalized Learning System (AI-PLS) for preschool education that integrates adaptive learning, multimodal interaction, and parental content customization.

The proposed system combines reinforcement learning, natural language processing, and collaborative filtering techniques to personalize educational content according to each child’s learning profile and real-time engagement signals. The AI-PLS architecture includes a learner profile engine, a reinforcement learning-based content sequencer using the Proximal Policy Optimization (PPO) algorithm, a repository of 480 interactive learning units, a multimodal interaction manager, and a parental dashboard supporting twelve configurable pedagogical dimensions. A ten-week experimental study involving 320 children aged 36–72 months from eight public kindergartens in Almaty and Astana, Kazakhstan, was conducted using intervention and control groups. Learning outcomes were evaluated using adapted BDI-2 subscales.

The results demonstrated statistically significant improvements in early literacy (23.4%), numeracy (17.8%), and sustained attention (19.1%) among children using AI-PLS compared to the control group. The system achieved an engagement retention rate of 89.3%. In the algorithmic evaluation, AI-PLS outperformed rule-based and machine-learning baseline models in predicting successful learning-unit mastery. Parental feedback also indicated high satisfaction with the customization interface. The findings suggest that, under supervised preschool conditions, AI-driven personalization may support short-term improvements in measured literacy, numeracy, attention, and engagement outcomes while strengthening family-centered content customization.

АННОТАЦИЯ

Раннее детское образование представляет собой критически важный период развития, в течение которого закладываются когнитивные, языковые и социально-эмоциональные основы. Однако традиционные подходы к обучению в дошкольных учреждениях часто не учитывают различия в готовности к обучению, предпочтениях в обучении, моделях вовлеченности и культурном происхождении детей в возрасте от трех до шести лет. Цель данного исследования — разработать и оценить персонализированную систему обучения на основе искусственного интеллекта (AI-PLS) для дошкольного образования, которая интегрирует адаптивное обучение, мультимодальное взаимодействие и персонализацию контента родителями.

Предложенная система сочетает в себе методы обучения с подкреплением, обработки естественного языка и коллаборативной фильтрации для персонализации образовательного контента в соответствии с профилем обучения каждого ребенка и сигналами вовлеченности в реальном времени. Архитектура AI-PLS включает в себя механизм профилирования обучающегося, секвенсор контента на основе обучения с подкреплением с использованием алгоритма оптимизации проксимальной политики (PPO), хранилище из 480 интерактивных учебных модулей, менеджер мультимодального взаимодействия и панель управления для родителей, поддерживающую двенадцать настраиваемых педагогических параметров. Было проведено десятинедельное экспериментальное исследование с участием 320 детей в возрасте от 36 до 72 месяцев из восьми государственных детских садов в Алматы и Астане (Казахстан), с использованием экспериментальной и контрольной групп. Результаты обучения оценивались с помощью адаптированных субшкал BDI-2.

Результаты продемонстрировали статистически значимые улучшения в ранней грамотности (23,4%), счете (17,8%) и устойчивом внимании (19,1%) среди детей, использующих систему персонализированного обучения на основе ИИ, по сравнению с контрольной группой. Система достигла уровня удержания вовлеченности 89,3%. В рамках алгоритмической оценки AI-PLS превзошла rule-based и machine-learning baseline-модели в прогнозировании успешного освоения учебных модулей. Родительская обратная связь также показала высокий уровень удовлетворенности возможностями настройки контента. Полученные данные указывают на то, что в условиях контролируемого и педагогически сопровождаемого дошкольного обучения персонализация на основе ИИ может способствовать краткосрочному улучшению измеряемых показателей грамотности, счёта, внимания и вовлечённости, одновременно поддерживая при этом индивидуальную настройку образования, ориентированную на семью.

Keywords: preschool education, personalized learning, artificial intelligence, reinforcement learning, parental customization, early childhood development, adaptive systems, natural language processing, engagement, Kazakhstan.

Ключевые слова: дошкольное образование, персонализированное обучение, искусственный интеллект, обучение с подкреплением, родительская настройка, развитие детей раннего возраста, адаптивные системы, обработка естественного языка, вовлечённость, Казахстан.

Introduction

The preschool years, spanning approximately ages three to six, constitute a neurologically sensitive period during which synaptic density, language acquisition capacity, and foundational cognitive schemas are established at rates unmatched across the lifespan [1-3]. Educational interventions applied during this window yield substantially higher returns per unit investment compared to interventions targeting older learners, as demonstrated by longitudinal economic analyses of early childhood programs [4]. Yet the global provision of preschool education remains dominated by instructor-led, curriculum-uniform approaches that are structurally incapable of responding to the developmental heterogeneity inherent in any cohort of preschool-aged children [5].

Technology-mediated learning environments have proliferated in early childhood contexts over the past decade, ranging from tablet-based educational games to interactive storytelling platforms [6]. However, the majority of these systems operate on fixed content libraries delivered in a predetermined sequence, offering minimal adaptation to individual learner characteristics [7]. The absence of principled personalization mechanisms means that cognitively advanced children encounter insufficiently challenging material, while children developing at a slower pace face content that exceeds their proximal zone of development—both scenarios resulting in suboptimal learning outcomes and disengagement [8].

Parental involvement is a well-established predictor of preschool learning success. Studies consistently demonstrate that children whose parents actively participate in educational content selection and review exhibit stronger literacy and numeracy outcomes than peers without such parental engagement [9]. Nevertheless, existing digital learning platforms provide caregivers with only superficial control—typically limited to enabling or disabling broad content categories—rather than fine-grained customization aligned with the child's cultural context, linguistic background, and the family's pedagogical priorities [10-11].

This paper addresses these gaps through the design, implementation, and evaluation of an AI-based Personalized Learning System (AI-PLS) for preschool education. The primary contributions of this work are: (i) a reinforcement learning-driven adaptive engine that continuously updates child-specific learning trajectories based on engagement and performance signals; (ii) a natural language processing module enabling age-appropriate voice interaction in Kazakh, Russian, and English; (iii) a twelve-axis parental customization dashboard providing granular control over content, pacing, and pedagogical emphasis; and (iv) a ten-week field evaluation across eight Kazakhstani kindergartens demonstrating statistically significant improvements in early literacy, numeracy, and sustained attention relative to a matched control group.

Materials and methods

The AI-PLS was developed over eighteen months through an iterative design process involving three principal phases: requirements elicitation, system architecture development, and empirical validation. Requirements were gathered through structured interviews with twenty-two preschool educators, fourteen child development specialists, and thirty-eight parents of preschool-aged children across Almaty, Astana, and Shymkent. Thematic analysis of interview transcripts identified eighteen functional requirements and nine non-functional requirements that were subsequently prioritized using a modified MoSCoW framework.

The study employed a controlled field-study design with independent pre-test/post-test assessment and cluster-level randomization. The experimental design compared preschool children using the AI-PLS platform with children following the standard kindergarten curriculum without access to AI-PLS. Cluster-level allocation was selected to minimize contamination effects that could arise if children within the same kindergarten environment were assigned to different intervention conditions.

Participants were recruited from eight public kindergartens located in Almaty and Astana. Institutional approval for participation was obtained from kindergarten administrations before participant enrollment began. Parents or legal guardians of eligible children received detailed information regarding the study objectives, procedures, and data protection measures prior to participation. Inclusion criteria required children to be between 36 and 72 months of age, regularly enrolled in preschool education, and without severe developmental disorders requiring specialized intervention. Children with prolonged absence rates exceeding 20% during the study period were excluded from the final analysis. Prior to participation, institutional agreements were established with all participating kindergartens, after which written informed consent was obtained from parents or legal guardians for all enrolled children.

Ethical approval and study documentation were obtained before the beginning of the intervention. The study protocol included institutional permission from participating kindergartens, parental informed consent forms, anonymized participant coding procedures, assessment forms, and data-retention rules. All evaluation records were stored in anonymized form and were used only for aggregated statistical analysis. No personally identifiable information, photographs, raw voice recordings, phone numbers, addresses, or institutional group names were included in the analytical dataset or model-training data.

A total of 320 preschool children participated in the study. Four kindergartens were assigned to the intervention condition and four to the control condition using cluster-level random allocation in order to minimize contamination effects between groups. The intervention group consisted of 164 children, while the control group included 156 children.

Randomization was conducted at the kindergarten level to minimize contamination effects between groups. Baseline demographic analysis confirmed no statistically significant differences between the intervention and control groups in age distribution, gender composition, or initial BDI-2 scores. Before implementation of the intervention, all participants completed independent pre-test developmental assessments using selected Battelle Developmental Inventory, Second Edition (BDI-2) subscales administered by trained preschool educators and child development specialists who were not involved in the system development process.

Children in the AI-PLS group used the system during supervised 20-minute learning sessions over the ten-week period. Sessions were conducted daily under the supervision of preschool educators and trained research assistants to ensure consistent implementation conditions across institutions. The system was used as a supplementary educational activity and did not replace regular preschool instruction. Children in the control group continued the standard preschool curriculum without access to AI-PLS. This design allowed comparison between standard instruction and AI-supported personalized learning under similar institutional conditions.

Following completion of the ten-week intervention period, all children underwent post-test assessment using the same BDI-2 developmental subscales applied during baseline evaluation. Learning gains were calculated by comparing pre-test and post-test scores between intervention and control groups. In addition to developmental outcomes, engagement retention, parental satisfaction, and platform interaction metrics were analyzed to evaluate overall system effectiveness.

All participant data were anonymized prior to analysis using randomly generated identifier codes. No personally identifiable information was stored within the reinforcement learning training datasets. Written informed consent was obtained from parents or legal guardians before participation in the study. Data transmission between client devices and the cloud infrastructure was protected using AES-256 encryption and secure HTTPS communication protocols.

To preserve participant privacy while maintaining research verifiability, all children were assigned randomly generated anonymous identifiers. The linkage between participant identity and research identifier was not used in statistical analysis and was not included in the model training dataset. Only aggregated demographic characteristics, anonymized pre-test and post-test scores, attendance rates, parental customization variables, and session-level interaction logs were used for evaluation. Assessment sheets were completed by independent preschool educators and child development specialists using anonymous child identifiers and standardized scoring procedures applied equally across both groups. The assessors were not involved in the development of AI-PLS.

Examples of collected anonymized variables included child identifier code, intervention/control assignment, age in months, gender, baseline and post-intervention BDI-2 literacy scores, numeracy scores, attention scores, attendance rates, session duration, task completion statistics, engagement metrics, and parental customization settings. Personally identifiable information including names, phone numbers, addresses, institutional group names, photographs, and raw voice recordings were excluded from analytical datasets and were not used for statistical analysis or reinforcement learning model training.

Parents retained the right to withdraw their child from the study at any stage. Access to anonymized datasets was restricted to the research team. Raw voice recordings and directly identifying information were not used for model training or statistical analysis. Data retention and deletion procedures followed institutional data protection requirements.

To support research transparency, the study protocol included institutional permission letters from participating kindergartens, parental informed consent forms, anonymized participant coding sheets, pre-test and post-test assessment forms, and aggregated statistical datasets. Due to child-data protection requirements, personally identifiable data and raw interaction records cannot be publicly disclosed; however, anonymized aggregated data, statistical tables, and the evaluation protocol may be made available from the authors upon reasonable academic request. This verification procedure was intended to balance research reproducibility with the privacy requirements of studies involving preschool children.

Table I summarizes baseline characteristics of the study population. As shown, no statistically significant differences were found between the intervention and control groups prior to implementation of AI-PLS (p > 0.05 for all variables), confirming successful random allocation and ensuring internal validity of the experimental design.

Table 1.

Baseline Demographic and Pre-Intervention Characteristics

Compared to existing adaptive preschool learning systems [1-3], AI-PLS demonstrated superior engagement retention and broader parental customization functionality, particularly in multilingual and culturally adaptive educational scenarios. Unlike rule-based adaptive systems, the proposed reinforcement learning framework continuously updates child-specific learning trajectories based on longitudinal interaction patterns.

The system architecture comprises five interconnected modules: (1) a learner profile engine maintaining dynamic developmental state vectors; (2) a reinforcement learning-based content sequencer implementing a Proximal Policy Optimization (PPO) algorithm; (3) a content repository of 480 interactive units spanning six developmental domains; (4) a multimodal interaction manager supporting voice, touch, and gesture inputs; and (5) a parental configuration dashboard with real-time synchronization. The system was deployed on Android tablets (8-inch screens) with a backend hosted on Kazakhstani government-approved cloud infrastructure. The system architecture complied with national data protection requirements applicable to educational technologies deployed in Kazakhstan.

The reinforcement learning agent treats content delivery as a Markov Decision Process where the state space encodes a 32-dimensional child profile vector (engagement metrics, recent performance, session history, fatigue indicators), the action space comprises 480 content selection decisions, and the reward function integrates three signals: task completion rate (weight 0.4), engagement duration (weight 0.35), and assessed learning gain (weight 0.25). Before the main evaluation, the agent was initialized using anonymized prototype interaction logs collected during the iterative development and pilot-testing phase involving 85 child profiles.

Table 2.

AI-PLS System Design Parameters and Specifications

The parental customization interface provides control across twelve content axes: (1) primary instruction language, (2) secondary language exposure level, (3) cultural theme emphasis (Kazakh, Russian, international), (4) daily screen-time ceiling, (5) session break frequency, (6) cognitive challenge gradient, (7) social-emotional content density, (8) motor skill activity inclusion, (9) religious sensitivity filters, (10) gender-neutral content preference, (11) topic avoidance lists (user-defined), and (12) reward animation style. Parameter changes propagate to the content sequencer within five seconds via a REST API with WebSocket update channels, ensuring that parental preferences are reflected from the child's next interaction.

Statistical analysis was conducted using IBM SPSS Statistics 28.0. Between-group differences in learning outcome gains were assessed using independent-samples t-tests following confirmation of normality via Shapiro-Wilk tests. Effect sizes were computed using Cohen's d. Engagement retention was defined as the proportion of enrolled children completing at least 70% of scheduled sessions. Pre-test and post-test gain scores were compared between intervention and control groups to evaluate the educational impact of AI-PLS over the ten-week study period. Comparative algorithmic evaluation was conducted against four baseline approaches: a rule-based recommendation baseline, a decision tree classifier, a random forest classifier, and an LSTM neural network model. These baselines were evaluated on the same held-out labeled interaction dataset to compare the ability of different models to predict successful mastery of recommended learning units. Normality assumptions were verified using the Shapiro–Wilk test prior to parametric analysis. Effect sizes were interpreted according to Cohen’s standardized criteria for educational intervention studies. The Battelle Developmental Inventory, Second Edition (BDI-2), has been widely validated for assessing early childhood developmental outcomes and was adapted for the Kazakhstani preschool context.

The PPO reinforcement learning model was trained using a learning rate of 0.0003, discount factor (γ) of 0.99, clipping parameter of 0.2, batch size of 64, and entropy regularization coefficient of 0.01. To reduce overfitting, the system employed dropout regularization, early stopping based on validation reward convergence, randomized session sampling, and cross-validation using temporally separated interaction datasets. The held-out validation cohort was separated from the reinforcement-learning pre-training data and was used only for model-performance evaluation.

To ensure independent validation of the proposed system, performance evaluation was conducted using a held-out assessment cohort separate from the reinforcement learning pre-training dataset. All pre-test and post-test assessments were administered by preschool educators and child development specialists who were not involved in the system development process. Statistical significance was evaluated using independent-samples t-tests with effect size estimation through Cohen’s d to verify the robustness of observed learning gains.

Results and Discussions

Table III presents the comparative performance of AI-PLS against four benchmark systems on a standardized evaluation protocol using a held-out assessment cohort (n = 96 children). The proposed AI-PLS achieves an overall learning outcome accuracy of 91.7% and an AUC-ROC of 0.951, representing a 20.3 percentage point improvement over the rule-based baseline and a 6.1 percentage point improvement over the LSTM neural network comparator. The difference between AI-PLS and the LSTM comparator is statistically significant (paired t-test: t(95) = 3.84, p < 0.001, d = 0.71). The held-out cohort included 96 children and produced 480 labeled interaction records used for classification evaluation. These records were used only for algorithmic validation and were not included in the PPO pre-training dataset. For the classification evaluation, the target variable was defined as successful mastery of a learning unit after interaction. A positive class represented successful mastery based on post-task assessment scores, while a negative class represented incomplete mastery or task failure. Accuracy, precision, recall, F1-score, and AUC-ROC were used to evaluate how reliably each model predicted whether a recommended learning unit would be successfully mastered by a child.

Table 3.

Classification Performance on Test Set (n = 480 labeled interaction records from 96 children)

Table 4.

Pre-Test and Post-Test Learning Outcome Scores by Group

Within the field evaluation cohort, children in the AI-PLS intervention group demonstrated a mean early literacy gain of 23.4 percentage points (SD = 4.2) compared to 9.1 percentage points (SD = 5.7) in the control group (t(318) = 21.4, p < 0.001, d = 2.38). Numeracy gains were 17.8 pp (SD = 3.9) versus 7.3 pp (SD = 4.4) in controls (t(318) = 18.6, p < 0.001, d = 2.07). Sustained attention, measured via the BDI-2 attention subscale, improved by 19.1 pp (SD = 5.1) in the AI-PLS group compared to 8.4 pp (SD = 5.8) in controls (t(318) = 14.3, p < 0.001, d = 1.95). The observed effect sizes indicate strong between-group differences in the measured developmental outcomes.

Engagement retention across the ten-week deployment reached 89.3%. This result indicates that the proposed system maintained a high level of child participation across the ten-week intervention period. However, because engagement retention depends on deployment context, future studies should compare AI-PLS with external preschool learning platforms under identical experimental conditions. Analysis of session-level engagement logs revealed that the reinforcement learning agent's content sequencing decisions correlated positively with sustained engagement (r = 0.67, p < 0.001), with the highest engagement scores associated with content units that the agent had selected within one standard deviation above the child's recent mean difficulty level—consistent with Vygotsky's zone of proximal development framework.

The parental customization interface was utilized actively by 84.7% of parents in the intervention group, with language selection and cognitive challenge gradient being the most frequently adjusted parameters (modified by 91.2% and 78.6% of active users, respectively). Parental satisfaction scores increased monotonically across the three assessment waves, reaching a mean of 4.61/5.00 (SD = 0.38) at week ten. Qualitative feedback highlighted the real-time synchronization feature and the transparency of the system's content rationale display as particularly valued affordances.

A subgroup analysis by age cohort revealed that children aged 48–60 months demonstrated the largest absolute learning gains under AI-PLS (early literacy: +26.1 pp), while the 36–48 month cohort showed the greatest relative effect size (d = 2.61 for numeracy). These findings suggest that the system's adaptive mechanisms are particularly effective for children in the middle preschool range, where developmental trajectories exhibit the greatest between-child variance. The 60–72 month subgroup showed robust gains but with a somewhat smaller effect size (d = 1.74), potentially reflecting ceiling effects on the BDI-2 subscales for this older cohort.

The NLP module achieved a child speech recognition accuracy of 94.2% in Kazakh and 96.1% in Russian across the evaluation period, exceeding the 88.5% threshold identified as necessary for reliable voice-based interaction with preschool-aged children in the requirements phase. Speech recognition accuracy was calculated from anonymized system-generated recognition logs and educator-verified interaction transcripts; raw voice recordings were not stored or used for model training. Accuracy was lower for the 36–42 month subgroup (88.7%) due to shorter mean utterance lengths and higher phonological variability, suggesting that continued training data collection for this developmental subpopulation would yield further performance gains.

Compared with prior AI-based personalized learning studies in early childhood education [1–8], the proposed AI-PLS demonstrates several methodological and practical advances. Existing preschool personalization systems have primarily focused on static rule-based adaptation [1], generalized AI-assisted assessment [2], or conceptual frameworks for personalized learning [3,5], while relatively few studies have evaluated reinforcement learning-driven adaptation using controlled field deployments in authentic preschool environments. Unlike previous approaches that relied primarily on fixed recommendation rules or teacher-configured sequencing, AI-PLS continuously adapts instructional trajectories through longitudinal interaction modeling and PPO-based policy optimization.

Furthermore, several existing studies in AI-supported preschool education have been limited by small pilot samples, laboratory-oriented evaluations, or the absence of control-group comparison designs [4,8]. In contrast, the present study employed a cluster-randomized controlled field design involving 320 preschool children across eight public kindergartens with independent pre-test/post-test assessment using standardized BDI-2 developmental measures. The inclusion of parental customization mechanisms and multilingual interaction support additionally distinguishes the proposed system from prior adaptive preschool platforms that focused primarily on learner-system interaction without substantial caregiver involvement [2,6].

The observed engagement retention rate of 89.3% also exceeds values commonly reported in adaptive educational technology literature, where sustained engagement over multi-week deployments remains a major challenge [6,7]. The integration of parental controls, culturally adaptive content, reinforcement learning-based personalization, and multimodal interaction therefore represents a more comprehensive preschool AI-learning framework than most currently reported systems.

Several limitations should be acknowledged when interpreting the results of this study. First, the evaluation period was limited to ten weeks; therefore, the findings should be interpreted as short-term learning outcomes rather than evidence of long-term developmental improvement. Although the observed gains in early literacy, numeracy, and sustained attention were statistically significant, further longitudinal research is required to determine whether these improvements persist after the intervention and transfer to broader preschool learning contexts.

Second, the study was conducted in eight public kindergartens in Almaty and Astana, Kazakhstan. As a result, the generalizability of the findings to other regions, private preschool institutions, rural settings, or children with different linguistic and socio-economic backgrounds remains limited. Future studies should include a more diverse sample across different educational environments.

Third, the use of AI-PLS may have introduced a novelty effect. Since children in the intervention group interacted with a new digital learning system, part of the observed engagement may be associated with the novelty of the technology rather than the personalization mechanism alone. Longer-term studies are needed to distinguish sustained pedagogical value from short-term motivational effects.

Fourth, although the system was designed to personalize content based on learner profiles and interaction signals, AI-based recommendation systems may introduce algorithmic bias if the training data do not sufficiently represent different developmental levels, languages, cultural backgrounds, or learning needs. To reduce this risk, the system incorporated parental content controls, educator supervision, and continuous monitoring of recommendation outputs. However, further validation with more diverse datasets is required.

Fifth, the implementation of AI in preschool education raises concerns regarding excessive screen exposure. In this study, AI-PLS was used only as a supplementary educational activity under supervised conditions and was not intended to replace traditional play, social interaction, or educator-led instruction. Future implementations should define strict session-duration limits and ensure that AI-based activities remain balanced with physical, social, and creative learning experiences.

Sixth, the collection and processing of child interaction data require strict privacy safeguards. Although the study used anonymized identifiers and did not include personally identifiable information in the analytical dataset, future deployments should maintain clear data retention policies, parental consent procedures, access control mechanisms, and transparent explanations of how learner data are used for personalization.

Finally, AI systems in preschool education should not be interpreted as substitutes for educators or parents. The proposed AI-PLS is intended to support personalized learning, provide adaptive content sequencing, and strengthen parental involvement, while final pedagogical responsibility should remain with human educators and caregivers. Therefore, the results of this study should be viewed as evidence of the potential of supervised AI-assisted preschool learning rather than as proof of autonomous AI-based instruction.

Several promising directions emerge from the findings of this study. First, future research should investigate the long-term developmental impact of AI-assisted personalized preschool learning through longitudinal follow-up studies extending into primary school education. Such studies would help determine whether improvements in literacy, numeracy, and sustained attention persist beyond the intervention period and transfer to later academic achievement.

Second, future system iterations should incorporate more advanced multimodal sensing approaches, including biometric and affect-aware engagement estimation using facial expression analysis, gaze tracking, speech prosody analysis, and physiological interaction indicators under strict privacy-preserving conditions. These additions may improve the precision of learner-state estimation and adaptive content sequencing.

Third, additional research is needed to evaluate the applicability of AI-PLS in more diverse educational contexts, including rural schools, multilingual environments outside Kazakhstan, private preschool institutions, and children with special educational needs. Extending the framework toward inclusive education settings may provide important opportunities for adaptive support of children with developmental delays, autism spectrum conditions, or language-learning difficulties.

Fourth, future studies should investigate collaborative human-AI educational models in which educators and AI systems jointly contribute to personalized learning decisions. Teacher-AI collaboration frameworks may improve transparency, pedagogical trust, and alignment between adaptive recommendations and classroom learning objectives.

Fifth, future work should further examine algorithmic fairness, explainability, and ethical governance mechanisms for AI systems in early childhood education. Although the current system incorporated parental controls and supervised deployment procedures, future implementations should include fairness auditing, interpretable recommendation explanations, and bias-detection protocols to ensure equitable educational outcomes across demographic and linguistic groups.

Finally, future technical development may explore federated learning architectures and on-device personalization methods to reduce centralized storage of child interaction data and further strengthen privacy protection in AI-assisted preschool learning environments.

Conclusion

This paper presented an AI-based Personalized Learning System for preschool education incorporating reinforcement learning-driven content adaptation, multimodal interaction, and a twelve-axis parental customization dashboard. In a supervised ten-week field study involving 320 children from eight Kazakhstani kindergartens, AI-PLS was associated with statistically significant short-term improvements in measured early literacy, numeracy, and sustained attention compared with the control group. The observed effect sizes indicate strong between-group differences in the measured outcomes, while the engagement retention rate of 89.3% shows that the system maintained consistent participation during the intervention period. Comparative evaluation against four algorithmic baselines showed that AI-PLS achieved higher prediction performance for learning-unit mastery than the baseline models. The parental customization interface, used actively by 84.7% of caregivers, demonstrates the potential of family-centered personalization in preschool educational technologies. Future work should include longer longitudinal follow-up, broader validation across diverse preschool settings, and further investigation of privacy-preserving and explainable AI mechanisms for early childhood education.

References:

1. Yekollu R. K. et al. AI-driven personalized learning paths: Enhancing education through adaptive systems //International Conference on Smart data intelligence. – Singapore : Springer Nature Singapore, 2024. – С. 507-517.
2. Güven G., Yumugan S. Artificial intelligence in preschool education:" Personalized assessment and evaluation with artificial intelligence" //Uluslararası Sosyal Bilimler ve Eğitim Dergisi. – 2025. – Т. 7. – №. 12. – С. 161-190.
3. Laak K. J., Aru J. AI and personalized learning //Educational Technology & Society. – 2025. – Т. 28. – №. 4. – С. 133-150.
4. Junruang C., Kanjug I. The Relationship between Computational and Creative Thinking in Preschool Children: An Application through Gamification and Artificial Intelligence-supported Constructivist Personalized Learning Environment //Educational Process: International Journal. – 2025. – Т. 17. – С. e2025337.
5. Kovalchuk V. et al. Artificial intelligence as an effective tool for personalized learning in modern education //Environment. Technology. Resources. Proceedings of the International Scientific and Practical Conference. – 2025. – Т. 3. – С. 187-194.
6. Murtaza M. et al. AI-based personalized e-learning systems: Issues, challenges, and solutions //IEEE access. – 2022. – Т. 10. – С. 81323-81342.
7. Fiza N. E. Z. et al. Navigating The Landscape of AI-Powered Personalized Learning in Primary Education: A Systematic Literature Review //Elementary Education: Journal of Studies, Analysis, and Development. – 2026. – Т. 1. – №. 1. – С. 1-17.
8. Oshinowo O. R. et al. An AI-Driven Teaching Prototype for Early Childhood Education: A Developmental Approach to Educational Management //Acitya Wisesa: Journal of Multidisciplinary Research. – 2025. – С. 45-65.
9. Yelland N., Lee L., O'Rourke M. Rethinking learning in early childhood education. – McGraw-Hill Education (UK), 2008.
10. Sun C. et al. Teacher-AI collaboration in content recommendation for digital personalised learning among pre-primary learners in Kenya //Proceedings of the Eleventh ACM Conference on Learning@ Scale. – 2024. – С. 346-350.
11. Gadzhimusieva D. et al. Development and pilot evaluation of an AI-driven learning management system for personalized education for autistic students //Education and Information Technologies. – 2026. – С. 1-39.