A Systematic Review of Deep Knowledge Tracing (2015-2025): Toward Responsible AI for Education

1. Introduction

The growing integration of Artificial Intelligence (AI) in education offers significant opportunities to enhance teaching practices, enrich student learning experiences, and streamline educational management. AI shows potential at both the classroom and system levels—for instance, by personalizing learning for diverse students, identifying those at risk, and assessing emerging competencies [1,2,3,4,5]. Central to most AI systems is the learner model—a computational representation of a student’s cognitive and non-cognitive characteristics. Learner models are often built from digital interaction data (e.g., interactions with learning systems) and provide real-time assessments of learner progress, allowing educators and AI tools to adjust instructions accordingly [6,7].

Two primary AI approaches are used to construct learner models: symbolic and sub-symbolic. Symbolic methods—such as rule-based systems and Bayesian networks—offer interpretable models and support the integration of expert knowledge, but they face challenges with noisy data and the high cost of encoding real-world educational problems into symbolic representations, a limitation known as the knowledge acquisition bottleneck [8]. In contrast, sub-symbolic methods—particularly deep learning models—can automatically learn complex, non-linear patterns from (non)sequential data, reducing reliance on manual feature engineering and enabling more scalable, data-driven learner modeling [9,10].

Deep Knowledge Tracing (DKT) is a prominent sub-symbolic approach that applies recurrent neural networks to capture the temporal dynamics of student knowledge [11]. Since its introduction, DKT has been shown to outperform traditional models in predictive accuracy, yet its advantages come with challenges—particularly opacity, potential bias, and a lack of pedagogical grounding [12,13,14,15,16]. In education, these shortcomings risk reinforcing inequalities and limiting meaningful adaptation to individual learner needs [3,17].

Addressing such issues requires moving beyond purely technical optimization toward responsible AI—a human-centered approach grounded in fairness, transparency, accountability, and educational validity [14,18,19]. As education is increasingly recognized as a high-risk domain for AI deployment (European Union’s AI Act [20]), trustworthiness and interpretability become not just technical goals but ethical imperatives for learner modeling. The next sections examine the evolution from classical to deep knowledge tracing, key deployment challenges from a responsible AI perspective, existing reviews, and the need for this study—outlining both the advances achieved and the challenges that remain—followed by the research questions.

1.1. From Classical to Deep Knowledge Tracing

To understand knowledge tracing in the context of learner modeling, it is essential to first clarify what phenomenon and process we are modeling. At the core of this approach lies the concept of knowledge components (KCs)—the fundamental cognitive units representing specific skills, concepts, or competencies that students must acquire and master through practice [21]. These KCs can range from basic arithmetic operations such as addition or subtraction to more complex problem-solving strategies in domains like physics, programming, or language learning. In educational environments, students engage with learning items such as exercises, questions, or problems, each of which requires the application of one or more KCs. As illustrated in Figure 1, the knowledge tracing process involves observing sequences of student interactions with learning items, where each interaction yields an observable response—typically coded as correct or incorrect. However, the actual knowledge state—the degree to which a student has mastered each KC—remains latent and must be inferred from these observable interaction patterns. The central computational challenge in KT is to model how this hidden knowledge state evolves over time as students practice, learn, forget, or consolidate their understanding, and subsequently to predict future performance on new or unseen items based on historical interaction sequences [21,22]. This modeling task is inherently sequential because learning is a temporal process where prior experiences influence current knowledge states, which in turn determine the probability of success on subsequent learning opportunities.

Knowledge Tracing (KT)—one of the most extensively studied approaches for learner modeling—estimates how a learner’s mastery of latent knowledge components evolves over a sequence of practice opportunities. Classical KT methods have been foundational in educational data mining since the 1990s. Bayesian Knowledge Tracing (BKT), introduced by Corbett and Anderson [23], models student learning as a Hidden Markov Model with binary latent states (mastered/not mastered) and four parameters: initial knowledge, learning rate, guess, and slip probabilities. Its strengths include strong theoretical grounding in cognitive science, interpretable parameters aligned with educational theories, and computational efficiency. However, the original BKT formulation suffers from several limitations: the binary mastery assumption oversimplifies continuous learning processes [24,25], the standard model does not incorporate forgetting nor handle multiple skills simultaneously [23,26], and parameter identifiability issues can lead to degenerate solutions where different parameter combinations yield identical predictions [24,27,28]. While numerous extensions have been developed to address these limitations, including forgetting-aware variants and multi-skill formulations, these enhancements often increase model complexity and computational requirements (e.g., [2,29]).

Performance Factor Analysis (PFA) by Pavlik et al. [29] addresses some BKT limitations by using logistic regression to predict performance based on prior successes and failures, allowing for more granular skill modeling. While PFA offers greater flexibility and can handle multiple skills, it still relies on hand-crafted features, assumes linear relationships between practice and performance, and cannot capture complex temporal patterns or student-specific learning trajectories [9,30]. Other classical approaches like Additive Factor Model (AFM) and Item Response Theory (IRT) variants face similar trade-offs between interpretability and expressiveness, often requiring extensive domain expertise for feature engineering and struggling to capture the non-linear, heterogeneous nature of real student learning [31,32,33]. These limitations stem from the need for sophisticated parameter tuning and feature selection processes that demand substantial educational domain knowledge [34], while their linear assumptions fail to adequately model the complex cognitive processes and heterogeneous learning patterns observed in real educational contexts [35,36].

While early work (e.g. BKT, PFA) relied on hand-crafted features and symbolic approaches [9,37], the introduction of DKT by Piech et al. [11] marked a major change by applying recurrent neural networks (specifically LSTMs) to capture complex temporal dependencies in student learning. The term "deep" refers to the use of multi-layer neural networks, and the original DKT specifically employed deep recurrent architectures (LSTMs). Since knowledge tracing is fundamentally a temporal prediction task, numerous other deep learning approaches beyond recurrent architectures have been developed for time series forecasting, including transformers, graph neural networks, and convolutional architectures [38,39,40]. Notably, transformers avoid recurrence entirely, relying solely on attention mechanisms [41], and have been successfully adapted to knowledge tracing (as discussed later in the Attentive model category). This sub-symbolic approach promised to overcome the limitations of traditional methods through automatic feature learning and improved predictive accuracy [30,42]. Since its introduction, numerous studies have focused on improving DKT’s prediction accuracy and addressing its limitations [22,43]. However, despite its ability to predict, DKT has revealed significant methodological challenges. Yeung and Yeung [44] identified critical issues with consistency over time: the model sometimes predicts lower mastery after correct answers and produces sequentially inconsistent predictions that weaken the assumption of gradual knowledge acquisition. Xiong et al. [13] showed that DKT does not perform much better than simpler models like PFA on prepared datasets, while also revealing weaknesses in handling multi-skill sequences and raising concerns about dataset quality. Wang et al. [45] further analyzed DKT’s instability through finite-state automata, showing that prediction volatility stems from built-in structural limitations rather than merely noisy data.

Additionally, DKT faces challenges common to deep neural networks: limited ability to integrate prior domain knowledge, susceptibility to learning spurious correlations from biased data (e.g., imbalanced datasets), and lack of interpretability [14,46]. Subsequent variants, such as self-attentive or graph-based KT, have improved predictive accuracy [47,48]. Yet, these technical refinements have not fully resolved underlying issues related to stability and transparency. The persistence of these problems—despite a decade of research—highlights the need for a critical examination of DKT research.

1.2. Critical Problems in DKT Deployment from Responsible AI Perspective

The concept of responsible AI has been articulated through various frameworks, each emphasizing key principles such as fairness, privacy, accountability, and transparency. For instance, Maree et al. [49] highlight fairness, privacy, accountability, transparency, and soundness, while Arrieta et al. [50] expand this view to include ethics, security, and safety. Other perspectives, such as those of Eitel-Porter [51] and Werder et al. [52], stress explainability as a crucial component. A broader perspective is offered by Jakesch et al. [53], incorporating sustainability, inclusiveness, social good, human autonomy, and solidarity into the responsible AI discourse. Building on these foundations, Goellner et al. [54] define responsible AI as a human-centered approach that builds user trust through ethical decision-making, explainable outcomes, and privacy-preserving implementation.

Synthesizing these contributions, this review examines DKT through the lens of responsible AI, drawing on prior critical reviews and surveys that link DKT research to key principles found across responsible AI frameworks. These include dealing with data quality issues, ensuring sequential stability so that pedagogical decisions remain reliable over time, and transparency—operationalized through process-level interpretability to support educator trust and regulatory compliance. Poor data quality (e.g., data noise, imbalanced distributions, sparsity) can undermine fairness by producing biased or inequitable learner assessments that disadvantage certain student groups or learning trajectories [55,56,57]. Overlooking sequential stability undermines accountability by leading to inconsistent or unreliable knowledge predictions, potentially resulting in inappropriate instructional actions [14,15,58]. A lack of interpretability risks educator distrust and non-compliance with oversight standards, whereas clear, process-level explanations strengthen transparency and help justify AI-driven decisions [59]. Together, these dimensions operationalize fairness, accountability, and transparency in the deployment of DKT, forming the foundation for responsible and trustworthy use in educational settings.

1.2.1. Data Inconsistency and Class Imbalance

Educational logs are inherently imbalanced: some items are trivially easy while others are rarely answered correctly [55,60]. This imbalance can cause DKT models to overfit to dominant patterns in the data, weighting their predictions toward frequently occurring or easier items. As a result, the models may misrepresent proficiency for learners who primarily encounter less-represented or more challenging items, indirectly creating biased assessments across different groups of students. Evidence of this emerges when mainstream knowledge tracing models are re-evaluated on resampled, balanced test sets: they show significant performance drops, revealing a heavy reliance on answer distribution biases rather than genuine learning patterns [56]. This bias is particularly pronounced in minority skills—knowledge components with limited representation in training data or extreme success/failure rates. Models may incorrectly classify students as proficient in rarely practiced skills based solely on overall performance patterns, or fail to detect genuine learning progress in challenging skills, leading to inappropriate instructional decisions and inequitable educational outcomes.

1.2.2. Sequential Stability

Sequential stability refers to a model’s ability to produce consistent, educationally plausible trajectories of a learner’s knowledge state over time. This construct encompasses several related aspects: temporal consistency (maintaining coherent predictions across time steps), smoothness (avoiding abrupt oscillations in mastery estimates), and educational plausibility (producing trajectories that align with learning theory). The lack of sequential stability creates serious problems for practical tutoring systems, as educators require mastery estimates that change smoothly and plausibly over time to make informed instructional decisions [44,45]. Current DKT models exhibit "wavy transitions" and dramatic fluctuations in knowledge state predictions even when students provide consistent responses, sometimes predicting lower mastery after correct answers [14,61,62,63]. These unstable trajectories prevent educators from trusting the system’s assessments, undermine actionable feedback, and may indicate problematic over-fitting that renders models unreliable for real-world deployment [44].

1.2.3. Process-Level Interpretability

The black-box nature of deep learning models creates significant barriers to adoption in educational contexts where educators demand explanations for AI-driven assessments [59,64]. This lack of interpretability prevents educators from understanding why the system makes specific predictions, undermining trust and potentially leading to harmful automated decisions. The absence of interpretable models also creates regulatory compliance problems, as the European Union’s AI Act requires explainability for high-risk AI systems in education.

While explainable AI methods such as LIME and SHAP exist [5,65], many current DKT models overlook implementing any interpretability mechanisms, creating a fundamental mismatch between technological capability and practical educational needs. It is important to distinguish between interpretability and explainability, terms that are often used interchangeably in the literature but represent distinct approaches to model transparency [16,66]. Interpretability refers to the degree to which a human can understand the cause of a decision through the inherent design of the model—achieved through architectural choices that make the decision-making process inherently transparent [67]. Explainability, in contrast, often refers to the ability to provide post-hoc explanations for model decisions through external methods applied after training, such as attention visualization or feature attribution techniques [68,69,70]. This distinction becomes crucial in educational contexts where both the immediate comprehensibility of model behavior (interpretability) and the ability to justify specific decisions to stakeholders (explainability) are essential for practical deployment and regulatory compliance. However, recognizing that many DKT studies in the literature use these terms interchangeably, this review adopts a pragmatic approach by classifying studies based on the actual techniques employed rather than the terminology used. Because the conceptual frameworks scholars cite can ultimately influence the techniques they select, resolving this misalignment remains a critical issue; establishing a clear view of what is currently implemented provides the necessary groundwork.

Collectively, these problems create urgent challenges for educational practice that require systematic evaluation. Educators need reliable, fair, and interpretable models to make informed instructional decisions, yet current DKT research fails to ensure these fundamental requirements are met. This necessitates an organized synthesis that looks beyond typical features and characteristics of DKT studies and examines how they address existing challenges to achieve responsible AI principles.

1.3. Existing Reviews and the Need for This Study

Despite the growing interest in (deep)KT, existing reviews have predominantly concentrated on model taxonomies and performance comparisons, offering limited insight into the deeper methodological challenges that affect the real-world deployment of these models. For instance, Dai et al. [6] conducted a review of KT techniques from the perspectives of assumptions, data, and algorithms, highlighting that most models address only a subset of assumptions about knowledge components and cognitive processes, and predominantly rely on quiz data as input. They identified dynamic Bayesian networks, logistic regression, and deep learning as the main algorithmic approaches. Song et al. [22] discussed different aspects of KT models to identify their distinctions and better support research in the field. Their review provided a granular categorization of mainstream models, detailed analyses of techniques and technological solutions, and outlined potential future research directions in deep learning–based KT. Shen et al. [71] conducted a comprehensive survey of KT, categorizing three fundamental model types, reviewing their variants under stricter learning assumptions, and showcasing typical application scenarios. They also introduced two open-source libraries—EduData for dataset access and preprocessing, and EduKTM for unified model implementation—to support research and practice, and discussed future development directions in the field. Bai et al. [72] provided a comprehensive review of explainable knowledge tracing (xKT), introducing core concepts from both xAI and KT, and categorizing KT models into transparent and black-box types. They reviewed interpretability methods across ante-hoc, post-hoc, and other dimensions, highlighted the lack of robust evaluation methods, and demonstrated three xAI approaches on ASSISTment2009, offering insights for improving interpretability evaluation from an educational stakeholder perspective. Abdelrahman et al. [21] provide a survey of KT models, covering methods from early approaches to recent deep learning–based techniques, while highlighting theoretical aspects, benchmark dataset characteristics, key modeling differences, and current research gaps, as well as outlining possible future research and application directions.

While these syntheses have undoubtedly contributed to advancing the (deep) knowledge tracing field by organizing research, guiding model development, and documenting methodological trends, notable gaps remain. No existing work has provided a systematic review of DKT that comprehensively collects and analyzes all relevant studies from its inception to the present. Crucially, most reviews focus on methodological taxonomies and performance comparisons but overlook practical challenges—such as data quality issues, sequential instability, and lack of interpretability—that are critical for ensuring the responsible AI compliance and trustworthiness of DKT systems. The current landscape therefore lacks an evidence-based synthesis that interrogates not only which models perform well, but also how they operate under real-world constraints that matter to educators and learners. This study addresses these gaps by presenting the first systematic review of DKT to explicitly evaluate methodological rigor in alignment with responsible AI principles, providing a foundation for advancing both technical performance and ethical readiness in educational AI.

1.4. Research Questions

This study synthesizes the DKT literature from 2015–2025 to identify critical issues that undermine reliable deployment and practical utility in educational settings. To address these problems and guide responsible AI–aligned deployment, the research questions are:

RQ1 (Modeling landscape): Which datasets, model architectures, input features (learner state indicators and context features), and validation metrics have been employed in DKT research, and how do these features distribute across the current taxonomy?

RQ2 (Data inconsistency): How do studies prevent biased student assessments by recognizing and mitigating data inconsistency issues in their predictive modeling?

RQ3 (Sequential stability): How do studies ensure reliable pedagogical decisions by evaluating the sequential stability of their models in relation to the evolution of knowledge?

RQ4 (Process-level interpretability): How do studies enable educator trust and regulatory compliance by providing interpretability for their DKT models?

2. Method

To design and implement this systematic review, we followed the guidelines outlined by PRISMA’s framework [73].

2.1. Database and Keywords

We conducted a thorough search across five major scholarly databases to ensure that our review was as comprehensive and unbiased as possible. These databases include Web of Science, Scopus, ScienceDirect, ACM Digital Library, and IEEE Xplore. Additionally, we explored grey literature by examining the first ten pages of Google Scholar [74] and selected the most pertinent research articles based on our predefined criteria. Our method involved employing keyword combinations identified through existing related works. The core search terms were formulated as ("knowledge tracing") AND (deep* OR neural*) and used to search within the title, abstract, and keywords, adapted to each database’s syntax requirements.

2.2. Eligibility Criteria

We established specific criteria to filter relevant articles. These criteria were divided into inclusion and exclusion categories, with inclusion criteria set in advance to guide our initial search. These included criteria such as publications being in English, peer-reviewed, and dated from 2015 to 2025. Exclusion criteria, refined during screening, removed studies without empirical evaluation, limited to classical knowledge tracing methods, or lacking sufficient methodological detail. For a detailed breakdown of these criteria, see Table 1.

2.3. Study Selection Process

Figure 2 illustrates the comprehensive PRISMA flow diagram illustrating the sequential stages of study identification, screening, and final selection. The article selection process comprised three primary steps of identification, screening, and eligibility evaluation. After the search, we imported our search results into Zotero reference management software and conducted both automatic and manual duplicate searches to eliminate duplications. Records identified from databases (n = 958) and Google Scholar (n = 89) were combined, with 311 duplicates removed.

In the first screening phase, the primary researcher evaluated abstracts and titles against the eligibility criteria, assigning scores of 0 (exclude), 0.5 (unsure), or 1 (advance to full-text screening). Another researcher independently coded a randomly selected subset to verify consistency with the primary researcher’s assessments. They also independently re-evaluated all abstracts initially scored as 0 and provided independent ratings for those marked as 0.5. Uncertain cases progressed to the second screening phase, while the first screening excluded 542 records with discrepancies resolved through consensus discussion.

In the second screening phase, 194 full texts were assessed following the same methodological approach, with the primary researcher conducting initial evaluations while the another researcher independently reviewed a subset and verified all articles scored 0 or 0.5. Additional exclusion criteria were refined during this stage as needed. Of the 194 full-text articles assessed, 5 could not be retrieved due to access limitations (E2), and 105 papers were excluded for specific reasons: lacking validation or empirical studies (E1, n = 29), classical KT only (E3, n = 32), no novel contribution (E4, n = 20; referring to comparative or applied studies using existing DKT frameworks without introducing methodological or conceptual innovations), insufficient methodological detail (E5, n = 17), and poster papers (E2, n = 7). All discrepancies were resolved through discussion until consensus was reached, resulting in 84 studies included in the final synthesis.

2.4. Quality Appraisal

Quality assessment is a pivotal stage in systematic literature reviews to ensure reliability and mitigate bias in study findings. Following the methodologies from previous reviews and guidelines [75], this review employs a tailored set of five criteria that emphasize the significance of internal and external validity, as well as the relevance and appropriateness of the research methodology and its execution. The five quality appraisal criteria are:

(1) Are the objectives and the context of research clear, and well connected with DKT?

(2) Does the research adequately delineate the methodology (e.g., training dataset, the model architecture, and its hyper-parameters)?

(3) Does the research adequately delineate the experiments (e.g., examining sequential or temporal consistency, assessing interpretability, addressing data inconsistencies, or evaluating the model’s real-world applicability)?

(4) Is there congruence between methodology and methods used for data collection, analysis, and interpretation?

(5) Are the study contributions and limitations clearly stated?

• The quality assessment framework allocated points according to predefined criteria: a response of "Yes" indicated high quality (1 point), "To some extent" indicated medium quality (0.5 points), and "No" indicated low quality (0 points). The highest possible score was 5. Studies scoring 3.5 points or higher were classified as high quality, those between 1.5 and 3.5 points as moderate quality, and those scoring 1.5 points or less as low quality. This approach ensured that only studies of sufficient quality were included, thereby supporting valid and reliable conclusions. All included articles received scores above the low-quality threshold, with all 84 papers falling into the medium- or high-quality categories.

2.5. Information Extraction and Data Analysis

Extracted information covered publication details (author names, country, year, publication type, keywords), study characteristics (research objectives, theoretical frameworks, sample), technical specifications (study design, data sources, DKT model architecture, evaluation metrics), methodological rigor indicators (data quality handling, sequential stability assessment, interpretability techniques), and study outcomes (findings, challenges, and future directions). Two researchers independently extracted and categorized this information using a standardized data extraction form.

For the first research question, we applied a deductive content analysis approach informed by the taxonomy of Abdelrahman et al. [21], which we adapted to suit the scope and objectives of this review. First, we reconceptualized "Text-Aware" category as a cross-cutting input modality rather than a standalone architectural category, mapping models that leverage textual features on their dominant computational mechanism (e.g., attention-driven text models are classified as Attentive, graph-enhanced text models as Graph-Based). Second, we introduced the Hybrid/Meta category to accommodate models that integrate multiple architectural paradigms or employ meta-learning approaches, which were not explicitly categorized in the original taxonomy. To ensure clarity and mutual exclusivity, we assigned each model to exactly one primary category based on its most distinctive architectural characteristics. These categorizations were derived from architectural references in the methods sections or, when unspecified, inferred from detailed model descriptions and methodologies in the reviewed articles. In addition to architectures, we systematically coded datasets, input features (including learner state indicators and context features), and validation metrics, drawing on detailed descriptions of model designs and methodologies extracted from the reviewed articles.

To address the second to fourth research questions, we employed an inductive thematic analysis approach to identify patterns in how studies addressed methodological challenges. This process involved three main activities: (1) identifying explicit mentions of data quality issues and the mitigation strategies reported in each study (e.g., missing data, data noise, imbalanced distributions, sparsity, inconsistent timestamps, and other data inconsistencies), (2) cataloguing the methods used to evaluate sequential stability—temporal consistency and smoothness—of model predictions (e.g., visual inspection through heatmaps, prediction trajectory plots, and moving-average smoothing curves), and (3) documenting the interpretability techniques applied (e.g., feature importance analysis, attention weight visualization, rule extraction, counterfactual examples, and concept-based explanations). Emergent themes were then iteratively refined through constant comparison across studies, with any coding disagreements resolved through discussion until consensus was reached. Accordingly, we created tables and visualizations to synthesize findings.

3. Results

Table A1 in the Appendix provides a comprehensive summary of all 84 included studies with their key characteristics. As shown in Figure 3, the included articles originated from eight countries (based on first author’s affiliation), with China contributing the majority of publications (e.g., IDs 1, 4, 6, 12), followed by Japan (e.g., IDs 17, 20, 40), the USA (e.g., IDs 14, 63, 84), Australia (e.g., ID 16), Estonia (e.g., IDs 37, 80), Hong Kong (e.g., IDs 44, 58), the Republic of Korea (e.g., IDs 23, 71), and Canada (e.g., ID 49). Although the initial pool covered more countries, the application of eligibility criteria narrowed the final sample to 84 articles from these eight, with representation across Asia, North America, and Europe highlighting DKT’s multi-regional recognition. Figure 4 presents the temporal evolution of DKT studies across model categories from 2015 to 2025, highlighting rapid growth after the model’s introduction in 2015, with the highest peak in 2024 (37 papers) driven mainly by hybrid/meta (e.g., IDs 22, 31, 79) and graph-based models (e.g., IDs 26, 29, 33). The data reveals distinct developmental phases: initial sequence-based explorations (2015; e.g., ID 84), diversification into attention and graph architectures (2018–2020; e.g., IDs 14, 40), and recent convergence toward hybrid approaches that combine multiple paradigms (2021–2025; e.g., IDs 22, 31, 79). Finally, Figure 5 illustrates the bibliometric analysis of the included articles, with 33 conference papers (e.g., IDs 8, 10, 14) and 51 journal publications (e.g., IDs 1, 3, 17), respectively. The analysis reveals a significant increase in journal publications commencing in 2022 (e.g., IDs 2, 49, 69), a trend that aligns with the accelerated adoption of educational technologies during the global pandemic. Conference proceedings have maintained consistent publication rates throughout the review period (e.g., IDs 8, 10, 14), indicating sustained interest and ongoing methodological discourse within the research community.

3.1. RQ1: Modeling Landscape

The investigation of the current DKT modeling landscape used a four-part analytical framework examining: (1) taxonomic classification of model architectures and their evolution over time, (2) patterns of dataset utilization across the research collection, (3) characteristics of learners and their learning processes, modeled through behavioral indicators and contextual features, and (4) the evaluation metrics employed for model assessment. Application of this taxonomic classification revealed distinct architectural paradigms with varying prevalence in the literature.

Hybrid/Meta models represent a large segment of the reviewed studies, accounting for 23.8% of the literature (e.g., IDs 10, 22), combining multiple paradigms or exploring novel architectures without a single dominant characteristic. Graph-based models comprise 26.2% of the studies (e.g., IDs 4, 6), explicitly capturing domain structure through graph neural networks, enabling the modeling of prerequisite relationships and skill dependencies that sequential approaches cannot address, in line with cognitive theories emphasizing the interconnectedness of knowledge [76]. Attentive models account for 17.9% of the studies (e.g., IDs 2, 5), employing self-attention mechanisms as the primary temporal modeling component, with categorization requiring that attention weights—not recurrent states—perform most of the temporal modeling. Sequence Modeling models represent 11.9% of the studies (e.g., IDs 8, 84) and rely on recurrent architectures without sophisticated attention or external memory, reflecting the foundational DKT paradigm through vanilla RNN/LSTM/GRU cores. Forgetting-aware models account for 9.5% (e.g., IDs 16, 77), explicitly incorporating temporal decay mechanisms to reflect cognitive theories of knowledge degradation, requiring mechanisms that reduce past knowledge contributions with elapsed time. Memory-augmented models represent 10.7% of the literature (e.g., IDs 25, 44), maintaining explicit external memory structures—key-value or slot-based—for learnable skill representations, which qualitatively differ from implicit hidden state propagation.

Figure 4 illustrates the temporal evolution of these categories from 2015 to 2025, showing the recent surge in DKT research and the current diverse ecosystem. The data reveals the field’s rapid growth, with hybrid approaches demonstrating increasing sophistication and graph-based approaches emerging prominently around 2019 in recent years, reflecting the field’s maturation and multi-paradigm value recognition.

DKT Data sources. Dataset utilization patterns across the reviewed corpus reveal dominance of ASSIST datasets, employed in 82.1% of the reviewed studies (69 of 84) (e.g., IDs 8, 12, 23), establishing a de facto standard benchmark while simultaneously introducing potential limitations in generalizability and cross-cultural applicability. This concentration reflects both pragmatic considerations—standardized benchmarks facilitate model comparison—and path dependencies that may constrain innovation in educational data representation.

The ASSIST family of datasets (ASSISTments 2009, 2012, 2015, and 2017) widespread adoption stems from several factors: public availability without restrictive licensing, pre-processed formats requiring minimal data engineering, established baseline performances enabling direct comparison, and comprehensive documentation of problem-skill mappings (e.g., IDs 8, 17). However, this ubiquity creates methodological monoculture where models optimized for ASSIST’s specific characteristics—binary correctness labels, single-skill assumptions, and particular student demographics—may fail to generalize to diverse educational contexts. EdNet, utilized in 26.2% of studies (e.g., IDs 5, 62, 71), introduces complexities absent in ASSIST: multi-modal question types, hierarchical content structures, and detailed timestamp granularity enabling fine-grained temporal analysis. Studies employing EdNet demonstrate increased attention to computational efficiency and scalability challenges, though cultural specificity in content and learning patterns limits global applicability. Statics2011, appearing in 14.3% of studies (e.g., IDs 23, 60), emphasizes multi-step problem solving and partial credit scoring that challenge binary classification assumptions prevalent in simpler benchmarks. Junyi, appearing in 8.3% of studies (e.g., IDs 20, 65), contributes large-scale K–12 activity logs commonly used to evaluate generalization beyond ASSIST. Programming judge datasets appear in 2.4% of studies (e.g., IDs 19, 56), highlighting domain-specific challenges such as compilation feedback, iterative debugging, and non-binary outcomes that differ from mathematics tutoring contexts.

The remaining studies employ diverse data sources including game-based learning environments (1.2%; e.g., ID 37) and synthetic datasets (2.4%; e.g., IDs 77, 84). Game-based datasets introduce engagement metrics and voluntary participation dynamics absent in classroom settings, while synthetic data enables controlled experimentation with known ground truth but risks oversimplifying real-world complexity. Notably absent from the corpus are datasets from developing educational contexts, vocational training, or adult learning environments, revealing geographic and demographic blind spots in current DKT research.

This dataset concentration creates cascading effects throughout the research pipeline. Studies exclusively using ASSIST demonstrate lower rates of addressing data quality issues (38%) compared to those employing diverse datasets (47%), suggesting that benchmark familiarity may breed complacency regarding data limitations. Furthermore, the predominance of mathematics and STEM content in available datasets potentially biases architectural innovations toward quantitative reasoning patterns while neglecting humanistic disciplines requiring different cognitive models.

• Modeled characteristics. Beyond dataset selection, the reviewed studies exhibit substantial variation in how they represent and model learner characteristics, as detailed in Table 2. Using the inductive analysis and labeling of the methodology section for each paper where authors described features used as inputs to their DKT models, we identified 1–3-word summaries whose prevalence is as follows: Task performance (the sequence of correct/incorrect answers) in 52 studies; Knowledge components/skills in 40; Extended interaction logs (raw learner interaction traces beyond basic sequences) in 17; Item difficulty in 16; Time-related features (response time or inter-event intervals) in 11; Textual features in 7; Behavioral signals (e.g., speed, attempts, hints, option choices, or media usage) in 4; and Knowledge mapping (mappings of items to skills or multiple skill requirements) in 2. The analysis reveals that most models rely on performance data, often augmented with a subset of temporal, semantic, or difficulty features.

Table 2. Input features used in DKT models.

Theme	Number of Articles	Article IDs
Task performance	52	1, 2, 5, 6, 7, 8, 10, 12, 13, 14, 15, 17, 20, 23, 24, 25, 26, 27, 30, 31, 32, 33, 34, 35, 38, 39, 40, 41, 44, 45, 46, 47, 49, 50, 53, 54, 59, 61, 62, 63, 65, 68, 69, 70, 71, 74, 75, 79, 81, 82, 83, 84
Knowledge components/skills	40	2, 3, 4, 9, 11, 13, 15, 16, 18, 21, 27, 28, 29, 30, 31, 35, 38, 39, 41, 43, 46, 47, 48, 49, 51, 52, 53, 56, 57, 58, 59, 62, 64, 66, 67, 72, 73, 77, 78, 82
Time-related features	11	2, 9, 11, 13, 16, 22, 43, 46, 64, 76, 80
Textual features	7	14, 18, 19, 54, 58, 60, 69
Item difficulty	16	7, 9, 10, 19, 27, 29, 34, 35, 43, 47, 48, 58, 60, 62, 66, 78
Behavioral signals	4	30, 61, 63, 73
Knowledge mapping	2	3, 4
Extended interaction logs	17	1, 3, 4, 7, 11, 22, 28, 36, 38, 40, 51, 52, 55, 67, 77, 78, 80

Table 2. Input features used in DKT models.

Theme	Number of Articles	Article IDs
Task performance	52	1, 2, 5, 6, 7, 8, 10, 12, 13, 14, 15, 17, 20, 23, 24, 25, 26, 27, 30, 31, 32, 33, 34, 35, 38, 39, 40, 41, 44, 45, 46, 47, 49, 50, 53, 54, 59, 61, 62, 63, 65, 68, 69, 70, 71, 74, 75, 79, 81, 82, 83, 84
Knowledge components/skills	40	2, 3, 4, 9, 11, 13, 15, 16, 18, 21, 27, 28, 29, 30, 31, 35, 38, 39, 41, 43, 46, 47, 48, 49, 51, 52, 53, 56, 57, 58, 59, 62, 64, 66, 67, 72, 73, 77, 78, 82
Time-related features	11	2, 9, 11, 13, 16, 22, 43, 46, 64, 76, 80
Textual features	7	14, 18, 19, 54, 58, 60, 69
Item difficulty	16	7, 9, 10, 19, 27, 29, 34, 35, 43, 47, 48, 58, 60, 62, 66, 78
Behavioral signals	4	30, 61, 63, 73
Knowledge mapping	2	3, 4
Extended interaction logs	17	1, 3, 4, 7, 11, 22, 28, 36, 38, 40, 51, 52, 55, 67, 77, 78, 80

• Evaluation metrics. Our analysis reveals broad agreement on evaluation metrics, with AUC being the most common measure, used in 76 studies (90.5%; all paper IDs except: 21, 42, 44, 48, 49, 55, 63, 72), making it the dominant standardfor assessing DKT model performance across diverse educational contexts. Accuracy was reported in 48 studies (57.14%; e.g., IDs 3, 10, 20, 22, 23) and Root Mean Square Error (RMSE) in 13 studies (15.48%; e.g., IDs 10, 29, 62, 72) as additional metrics. The heavy reliance on AUC reflects the field’s focus on binary classification and its consideration of imbalanced classes, though it does not fully capture the temporal dynamics and sequential dependencies that are central to student learning processes (see Section 3.3).

Table 3. DKT model categories with their core deep learning building blocks, representative models, and discriminative cues.

Category	Building-block sub-architecture	Representative KT models	Discriminative cue	Paper IDs
SequenceModeling	RNN / Long Short-Term Memory (LSTM) / Gated Recurrent Unit (GRU) layers	DKT [11], DKT+ variants	Hidden-state recurrence; no external memory or self-attention.	8, 20, 23, 28, 36, 52, 53, 60, 70, 84
Memory-Augmented	Key–value memory network	DKVMN, BCKVMN	External differentiable key–value slots updated each step.	25, 35, 44, 47, 48, 67, 68, 73, 78
Attentive	Pure self-attention / Transformer	SAKT, SAINT(+), AKT	Multi-head attention re-weights past items; no recurrence.	2, 3, 5, 7, 9, 11, 12, 14, 15, 18, 38, 56, 61, 74, 81
Graph-Based	Static concept-graph GCN	GKT, JKT	Learner state propagated on a fixed concept graph.	4, 6, 13, 26, 29, 33, 40, 41, 42, 43, 45, 46, 49, 50, 54, 55, 58, 59, 65, 66, 69, 72
Forgetting-Aware	Time-decay gating	LPKT, DKT-Forget, DGMN	Explicit decay functions/gates reduce distant influences.	16, 17, 24, 30, 51, 63, 76, 77
Hybrid/Meta	Ensemble or novel architectures	Mixing-Framework, SAKT+DKT, Graph+Decay	Combines multiple paradigms or explores novel architectures.	1, 10, 19, 21, 22, 27, 31, 32, 34, 37, 39, 57, 62, 64, 71, 75, 79, 80, 82, 83

Table 3. DKT model categories with their core deep learning building blocks, representative models, and discriminative cues.

Category	Building-block sub-architecture	Representative KT models	Discriminative cue	Paper IDs
SequenceModeling	RNN / Long Short-Term Memory (LSTM) / Gated Recurrent Unit (GRU) layers	DKT [11], DKT+ variants	Hidden-state recurrence; no external memory or self-attention.	8, 20, 23, 28, 36, 52, 53, 60, 70, 84
Memory-Augmented	Key–value memory network	DKVMN, BCKVMN	External differentiable key–value slots updated each step.	25, 35, 44, 47, 48, 67, 68, 73, 78
Attentive	Pure self-attention / Transformer	SAKT, SAINT(+), AKT	Multi-head attention re-weights past items; no recurrence.	2, 3, 5, 7, 9, 11, 12, 14, 15, 18, 38, 56, 61, 74, 81
Graph-Based	Static concept-graph GCN	GKT, JKT	Learner state propagated on a fixed concept graph.	4, 6, 13, 26, 29, 33, 40, 41, 42, 43, 45, 46, 49, 50, 54, 55, 58, 59, 65, 66, 69, 72
Forgetting-Aware	Time-decay gating	LPKT, DKT-Forget, DGMN	Explicit decay functions/gates reduce distant influences.	16, 17, 24, 30, 51, 63, 76, 77
Hybrid/Meta	Ensemble or novel architectures	Mixing-Framework, SAKT+DKT, Graph+Decay	Combines multiple paradigms or explores novel architectures.	1, 10, 19, 21, 22, 27, 31, 32, 34, 37, 39, 57, 62, 64, 71, 75, 79, 80, 82, 83

3.2. RQ2: Data Inconsistency and Bias

The organized examination of data quality issues (e.g., missing and noisy data, imbalanced distributions, sparsity, inconsistent timestamps, and other inconsistencies) across the 84 reviewed manuscripts reveals substantial shortcomings in how these fundamental concerns are methodologically addressed in contemporary DKT research.

Figure 6 presents a parallel sets diagram visualizing the relationships between data quality issues and their corresponding treatment techniques across the reviewed corpus. The majority of studies (56.0%) neglect data quality concerns despite their critical impact on model performance and generalizability. Across the corpus, sparse data is addressed in 22.6% (19/84) of studies, followed by imbalance (15.5% = 13/84) and Q-matrix bias (3.6% = 3/84). Additionally, isolated studies address causal de-bias (1.2% = 1/84) and data noise (1.2% = 1/84). The figure highlights the diversity of treatment techniques, ranging from embedding/cold-start methods and contrastive learning to preprocessing and graph/relational methods. The width of the connecting flows indicates how issue types map to multiple strategies, showing that researchers often adopt complementary approaches to manage complex data quality challenges. To further unpack these patterns, the following sections examine each category of data quality issues in detail, outlining the specific challenges they pose, and the strategies proposed to address them.

Sparse data (22.6% = 19/84 studies; e.g., IDs 11, 59, 71): This challenge arises when students exhibit limited interaction histories with specific skills or items, constraining reliable estimation of latent knowledge states. Reported mitigation strategies fall into four recurrent design patterns. First, cold-start initialization modules introduce learnable student and item (or skill) embeddings to enable prediction with minimal prior history (e.g., IDs 28, 73). Second, contrastive or comparative representation learning increases robustness under sparsity by pulling semantically or conceptually similar interaction pairs closer while pushing dissimilar ones apart, improving generalization from few observations (e.g., IDs 53, 73). Third, structural or relational propagation methods leverage auxiliary graphs—skill–skill, question–concept, or curriculum relations—to infer unobserved proficiency signals through neighborhood aggregation or counterfactual augmentation (e.g., IDs 41, 65, 69). Fourth, side-information integration enriches sparse sequences with additional modalities such as question text, difficulty or knowledge-structure references, and engineered statistics to densify representations (e.g., IDs 27, 59). Additional preprocessing tactics (e.g., interpolation, padding, sequence truncation, removal of ultra-short histories; IDs 37, 76) and sampling/negative sampling schemes (ID 4) complement these approaches by regularizing extremely short or irregular interaction traces. Collectively, these techniques operationalize sparsity handling through a blend of architectural inductive bias, representation learning, and data-centric augmentation rather than a single standardized procedure.

Imbalance (15.5% = 13/84 studies; e.g., IDs 9, 30, 36): This category combines approaches for handling skewed correctness distributions with data-centric augmentation. The problem occurs when certain skills are predominantly answered correctly (very easy) or incorrectly (very difficult), driving biased mastery estimates. Reported mitigations include: (1) contrastive or re-weighted representation learning modules that explicitly balance local and global interaction frequencies to keep minority trajectories salient (e.g., IDs 74, 75); (2) preprocessing heuristics such as filtering low-activity learners, truncating extreme frequencies, or normalizing heterogeneous features before model fitting (e.g., IDs 30, 44, 57); (3) imbalance-aware embeddings that adapt student–concept representations when population sizes differ markedly (e.g., ID 55); (4) relational propagation that leverages graph structure to smooth sparse curriculum regions and reduce majority-class dominance (e.g., ID 69); and (5) stratified sampling and autoencoder-based augmentation that manufacture additional minority interaction sequences (e.g., ID 80). Several studies simply acknowledge imbalance and rely on robust metrics such as AUC or balanced accuracy instead of explicit treatments (e.g., IDs 9, 38, 49, 51, 68), underscoring the field’s ongoing reliance on evaluation-side workarounds rather than holistic data rebalancing.

Data noise (1.2% = 1/84 studies; e.g., ID 82): Although rare, some work explicitly tackles logging artefacts and duplicated entries that distort learner histories. The representative study refreshes the ASSISTments release to merge duplicate interactions, discards incomplete records, and layers dropout regularization to suppress residual noise. These preprocessing pipelines emphasize the importance of curating interaction traces before model training when raw exports contain inconsistent or partially recorded events.

Q-Matrix bias (3.6% = 3/84 studies; e.g., IDs 3, 64, 72): Expert-defined Q-matrices, which specify the skills required for each question, often contain subjective biases or may be incomplete. To address this, studies have applied: (1) learnable skill embeddings, enabling the model to automatically discover skill–question relationships and override expert-defined mappings when they conflict with observed data (e.g., IDs 3, 64, 72); (2) automatic Q-matrix expansion, which identifies additional skill–question relationships not captured in the original expert specification (e.g., ID 64); and (3) skill–question relationship refinement, which adjusts the strength of expert-defined links based on empirical evidence (e.g., ID 72).

Causal de-bias (1.2% = 1/84 studies; e.g., ID 62): This emerging area focuses on algorithmic fairness and the mitigation of spurious correlations in knowledge tracing. The representative study integrates front-door adjustments into a causal self-attention mechanism to dampen confounding factors and recover more equitable mastery trajectories across demographic groups. Although nascent, this causal treatment illustrates how fairness-aware objectives can be embedded directly in the representation learning pipeline when data bias poses risks to student-facing predictions.

3.3. RQ3: Sequential Stability

Sequential stability refers to a model’s ability to generate smooth trajectories of a learner’s latent knowledge state as additional interactions are observed. This concept is important for three reasons. First, rapid oscillations in predicted mastery undermine the practical value of educational AI systems: teachers and students cannot rely on feedback that reverses direction after a single response, especially in contexts like K–12 adaptive tutoring where real-time instructional decisions are needed. Second, instabilities often indicate overfitting to unusual sequences and therefore predict poor generalization to unseen learners. Third, the cost of misclassification can vary across different points in a learning sequence: for example, mistakenly predicting mastery early on may prevent crucial practice opportunities, while a similar error near the end of instruction might carry less consequence, mainly influencing final assessment rather than ongoing learning. Despite its importance, sequential stability has received far less attention compared to commonly reported accuracy metrics such as AUC or RMSE.

Regarding sequential stability of predictions, our synthesis (Table 4) shows that qualitative visual inspection is by far the predominant method. Among the investigations reviewed, 62 of 84 (73.8%) present exemplar visualizations—typically heat maps of item-level probabilities or line charts of mastery trajectories—for a small number of learners. These graphical representations support model intuition but are typically applied to only a few test cases and are rarely paired with objective assessment criteria, offering limited evidence of reliability across larger learner populations. A smaller subset—3 of 84 (3.6%; IDs 29, 30, 77)—computes quantitative stability metrics, including the prediction-waviness score (sum of absolute first-order differences across trajectories), ${\ell }_2$-norm fluctuations between successive knowledge vectors, and the temporal-consistency coefficient originally introduced by Piech et al. [11]. These metrics capture different aspects of stability: waviness reflects smoothness, while temporal consistency reflects predictive coherence across time. Such measures enable direct model comparison but remain under-reported and lack established thresholds for acceptable stability [14].

Approximately 6.0% of studies (5 of 84) examine forgetting behavior, testing whether predicted proficiency decays during inactivity. This is useful for detecting models that unrealistically preserve full mastery indefinitely, though most analyses rely on synthetic rather than authentic gaps (e.g., IDs 2, 16, 19, 65). Only 1 paper extends evaluation to perturbation-based sensitivity analysis (1.2% of 84), injecting random noise into input sequences and measuring the resulting trajectory variance (ID 80).

Overall, no study in this corpus jointly reports stability metrics, forgetting, and sensitivity, indicating that evaluations of sequential stability remain fragmented rather than holistic.

3.4. RQ4: Process-Level Interpretability

Process-level interpretability refers to approaches that reveal how a knowledge tracing model generates predictions across a learning sequence, rather than focusing solely on end-point accuracy. In DKT, this entails clarifying the mechanisms, structures, or inputs that drive the evolution of a learner’s latent knowledge state and its observable outputs.

As summarized in Table 5, process-level interpretability remains underexplored in contemporary DKT research. A majority of studies (88.1%, 74 of 84) do not consider interpretability at all, offering no explicit strategies for explaining how predictions are derived. Among the remaining works that incorporate process-level interpretability, the most frequently implemented approach is feature attribution / attention analysis (9.5% of studies; e.g., IDs 1, 11, 31). These post-hoc techniques seek to explain a model’s current prediction by isolating which prior learner interactions exerted the greatest influence. Approaches include Deep SHAP (ID 11) for Shapley-based contribution scoring of historical interactions, attention-weight saliency inspection (e.g., IDs 31, 71) to highlight pedagogically salient exercises emphasized by Transformer or hierarchical attentive modules, and sequence-adapted gradient propagation methods such as Grad-CAM (ID 71) to spatially (temporally) localize decisive segments within interaction histories. An example of analytic feature saliency based on model sensitivity for this purpose was given in Linja et al. [70]. While these techniques provide granular, instance-level justifications, their interpretive value for educators is moderated by two limitations: (i) attention distributions or attribution scores do not always correspond to causal pedagogical factors, risking over-interpretation; and (ii) their effective use often presupposes technical literacy to contextualize attributions within curriculum design or assessment strategies.

The second most common architectural strategy is interpretability by design (1.2% of all studies; e.g., ID 60). These models embed educationally meaningful inductive structure directly into the architecture—such as skill- or concept-level parameterisation / factorisation layers or decomposable outputs—so intermediate variables can be audited. Such structural constraints increase semantic alignment with instructional constructs and can surface intermediate variables that are straightforward to audit. However, they often operate primarily as knowledge injection to enhance predictive calibration rather than fully disclosing internal state transitions or uncertainty propagation. Thus, transparency is partial: educators gain visibility into selected pedagogically grounded components but not necessarily into the full interaction of latent subsystems.

Finally, a small subset investigates embedding or trajectory visualization (1.2% of studies; ID 46), applying methods such as t-SNE projections of latent knowledge embeddings to reveal emergent clustering of skills or learner progression pathways. These visual artefacts offer intuitive, exploratory overviews of representational geometry but remain qualitative, lacking explicit alignment to decision-relevant thresholds or actionable formative feedback. They therefore complement but do not substitute for attribution- or design-oriented interpretability.

4. Discussion

This systematic review highlights notable advances in predictive learner modeling while also surfacing new insights into unresolved challenges that hinder the responsible deployment of DKT models in practice. In particular, it reveals a narrow modeling landscape dominated by a single dataset family, limited attention to data quality, and the near absence of sequential stability assessments. It also uncovers a critical transparency gap, as interpretability remains underdeveloped despite growing educational and regulatory demands. These findings together provide a novel, cross-cutting perspective on the systemic challenges that must be addressed for DKT to achieve sustainable impact.

4.1. Modeling Landscape and Architectural Evolution

The temporal evolution patterns of DKT architectures reveal a field that has progressed from early sequence modeling approaches to a diverse ecosystem dominated by Graph-Based models (26.2% of studies), followed by Hybrid/Meta (23.8%) and Attentive models (17.9%). This trajectory illustrates the field’s shift from simple sequential models toward architectures that promise richer relational reasoning and adaptive flexibility.

Dataset utilization patterns reveal a concerning over-reliance on ASSIST datasets, employed by 82.1% of studies. While these datasets provide valuable benchmarks, this concentration may perpetuate methodological blind spots and limit the generalizability of findings across diverse educational contexts [77]. The field’s implicit trust in established benchmarks potentially masks data quality issues that become apparent only when models encounter varied educational environments, as demonstrated by the finding that studies using diverse datasets show similar rates of data quality consideration (47%) compared to those relying solely on ASSIST (38%) [78].

Regarding the modeled characteristics, there is a notable trend towards incorporating more complex skill representations and relationships. Many studies (e.g., IDs 4, 6) leverage graph-based structures to model the interdependencies between skills, allowing for a more nuanced understanding of learner trajectories. This shift is indicative of a broader recognition of the importance of relational reasoning in educational contexts. Concerning the evaluation metrics employed for model assessment, the dominance of AUC (reported in 90.5% of studies) demonstrates near-universal adoption of a single discriminative benchmark. However, this heavy reliance reinforces a predominantly binary predictive framing that neglects sequential dependencies and temporal dynamics central to learning processes.

The implications of this architectural and geographical concentration extend beyond academic concerns. The dominance of certain research traditions and dataset preferences may inadvertently embed cultural and pedagogical assumptions into model architectures, potentially limiting their effectiveness when deployed in different educational contexts [60,77]. Future research must prioritize architectural designs that account for diverse learning environments and cultural contexts to ensure global applicability (Chinta et al. [79]).

4.2. Data Quality and Methodological Rigor

The critical evaluation of data quality treatment reveals one of the most concerning findings of this review: 56.0% of studies fail to acknowledge or address fundamental data quality issues inherent in educational datasets. This methodological limitation raises important concerns for the real-world deployment and educational validity of DKT systems, as it may contribute to biased assessments and reduce the reliability of recommendations [80,81].

The categorization of bias mitigation strategies demonstrates that when studies do address data quality, they predominantly focus on sparse data (22.6% of studies) and imbalance (15.5% of studies). These approaches often rely on architectural innovations rather than systematic data preprocessing, reflecting the sequential nature of knowledge tracing data where traditional resampling techniques risk disrupting temporal dependencies fundamental to accurate prediction [82].

The relationship between model architecture and data quality consideration shows a qualitative pattern: models that explicitly encode relational or structural information (e.g., graph-based or memory-augmented designs) more frequently describe data validation or augmentation steps than purely sequential baselines, likely because structural priors make inconsistencies in mappings and sparsity more visible [78,83].

The lack of systematic approaches to imbalance is particularly problematic given recent findings that mainstream knowledge tracing models show significant performance drops when evaluated on resampled, balanced test sets, revealing their heavy reliance on answer distribution biases rather than genuine learning patterns [56]. This suggests that the predictive accuracy reported in much of the existing literature may overestimate real-world performance and educational utility [30]. The emerging attention to causal de-bias (1.2% of studies) represents a critical but underexplored area. As educational AI systems directly impact students’ academic trajectories and future opportunities, ensuring algorithmic fairness across demographic groups becomes not merely a technical concern but an ethical imperative [18,79]. The field must urgently develop standardized protocols for bias detection and mitigation that account for the unique characteristics of educational data while ensuring equitable outcomes across diverse student populations [14,81].

4.3. Sequential Stability Assessment

The systematic evaluation of sequential stability of predictions reveals one of the most significant unresolved problem in current DKT research. Only 3.6% of studies employ quantitative sequential stability metrics, despite the critical importance of stable, educationally plausible knowledge trajectories for practical deployment in educational settings [44,45].

Variation across architectural paradigms is described but rarely systematically quantified in the source studies. Designs that explicitly encode temporal decay or memory structures more often report qualitative trajectory analyses, whereas pure attention-based architectures seldom introduce additional stability diagnostics beyond illustrative plots [62,78]. The implications of sequential instability extend beyond theoretical concerns. In practical educational settings, teachers and adaptive learning systems require models that produce stable, interpretable progressions of student knowledge. A model that achieves high predictive accuracy but exhibits "wavy transitions"—predicting mastery after one correct answer only to reverse this assessment after the next—provides little actionable guidance and may undermine educator trust in AI-driven insights [61,63].

The field’s evolution toward responsible AI in education necessitates fundamental reconsideration of evaluation practices. Future work must prioritize metrics that explicitly account for sequential dependencies and educational validity, including trajectory smoothness indices, learning progression coherence measures, and temporal consistency evaluations that ensure predicted knowledge sequences follow established developmental pathways [84,85].

4.4. Interpretability and Transparency

The analysis of interpretability mechanisms reveals a critical transparency deficit that directly conflicts with emerging regulatory frameworks and educational needs. Only 11.9% of studies provide any form of interpretability, with 88.1% deploying black-box architectures that offer no insight into their decision-making processes [59,64]. Model complexity shows an inverse relationship with interpretability. Memory-centric or factorised representations can support state inspection, whereas heterogeneous hybrid architectures rarely embed coherent explanatory interfaces [78]. This reflects a broader trade-off: as architectures grow more sophisticated to improve predictive accuracy, their internal reasoning often becomes more opaque [78].

In educational contexts, the distinction between interpretability and explainability is critical [67,68,86]. Most studies that attempt prediction explanation rely on post-hoc methods, such as SHAP or Grad-CAM, which may not faithfully capture the true decision-making process [16]. By contrast, interpretability-by-design approaches constrain models to remain transparent but may sacrifice predictive power [67,68,86]. This opacity creates both educational and regulatory risks. Teachers cannot meaningfully integrate AI-driven insights into practice without understanding the reasoning behind mastery predictions [59,64]. Under the European Union’s AI Act, deploying black-box models for student assessment may become legally non-compliant, creating both practical limitations and regulatory vulnerabilities for 88.1% of current models [87].

4.5. Cross-Cutting Patterns Across Responsible AI Dimensions

The synthesis across the three responsible AI dimensions—data quality, sequential stability, and interpretability—shown in Figure 7, reveals clear, category-specific patterns with direct implications for responsible deployment in education [14,18].

The first tension involves the conflict between predictive performance optimization and educational validity. Studies reporting the highest AUC scores (>0.85) rarely provide complementary interpretability or stability evidence, amplifying a performance–validity gap [72,88]. The field’s heavy reliance on AUC as the dominant metric (90.5% of studies) assumes independence of observations, a mismatch that risks incentivizing models which exploit statistical regularities rather than capturing authentic learning dynamics [89,90].

The second tension arises from dataset concentration effects. Over-reliance on ASSIST datasets (82.1% of studies) fosters implicit trust in established benchmarks while potentially obscuring data quality issues. In contrast, studies using more diverse datasets more frequently address methodological concerns across multiple dimensions, suggesting that benchmark diversity can function as a form of natural quality assurance [77,78].

4.6. Limitations

This review has inherent limitations that must be acknowledged. The English-only inclusion criterion may have excluded relevant research from non-English speaking countries. The taxonomic application required some subjective judgment despite clear criteria, and quantitative meta-analysis was not feasible due to heterogeneous reporting standards. While three researchers participated in this process to reduce bias, the classification may not be entirely free from subjectivity.

Our search strategy, though comprehensive across five major databases and grey literature, may also carry selection biases. The keyword combinations centered on "knowledge tracing" with ("deep*" OR "neural*") may have excluded studies using alternative terms such as "student modeling." Upon extensive meta-review of the existing works, we decided that DKT studies would reference these terms in their title, abstract, or keywords, though it remains possible that a very few studies might have been missed. The grey literature search via Google Scholar helped to mitigate this risk, although expanding the search further could potentially have identified additional relevant materials. Additionally, while this review covers diverse architectures including transformers, graph neural networks, and memory-augmented models, the original formulation of DKT focused on recurrent architectures, which shaped early field development and may have influenced the conceptual framing of knowledge tracing as primarily a recurrent modeling task. Finally, the review’s focus on empirical modeling studies means that theoretical contributions and methodological frameworks not directly tied to model development may be underrepresented.

5. Conclusions

This systematic review identifies both significant achievements and unresolved problems in DKT research from 2015–2025. While predictive accuracy has improved substantially through architectural innovations, the analysis reveals persistent problems that prevent practical deployment and compromise the educational validity of these models.

The most concerning finding is the widespread neglect of fundamental methodological requirements essential for responsible AI deployment in education. Only 3.6% of studies employ quantitative sequential stability assessment, despite its critical importance for generating educationally plausible knowledge trajectories. A majority (56.0%) fail to address data quality issues inherent in educational datasets, potentially compromising model reliability and fairness. Furthermore, 88.1% lack interpretability mechanisms, directly conflicting with emerging regulatory frameworks such as the EU AI Act, which mandates transparency for high-risk AI applications in education with full enforcement by August 2026.

The field’s over-reliance on AUC as the primary evaluation metric (90.5% of studies) reflects a fundamental methodological mismatch with the sequential nature of learning data. This emphasis on aggregate performance metrics may inadvertently incentivize models that exploit statistical regularities rather than capture genuine learning dynamics, undermining educational utility despite high reported accuracy. It also reinforces a binary framing of student knowledge, reducing inherently continuous and latent learning processes to discrete correctness outcomes. The geographic concentration of research and overwhelming dependence on ASSIST datasets (82.1% of studies) create potential blind spots that may limit the global applicability and cultural sensitivity of DKT systems. Studies utilizing more diverse datasets demonstrate higher rates of methodological rigor across multiple dimensions, suggesting that benchmark diversity can serve as a natural quality assurance mechanism. The synthesis reveals that methodological advancement has not kept pace with architectural innovation. Only a very small minority of studies address all four critical dimensions (modeling, data quality, sequential stability, and interpretability) with substantive depth. No single architectural category demonstrates consistent methodological strength across all dimensions, challenging the assumption that computational sophistication naturally leads to educational validity.

As AI systems become increasingly integrated into educational settings, the need for responsible AI frameworks encompassing transparency, fairness, accountability, and educational validity becomes paramount. The field has to evolve beyond accuracy-focused development toward educationally valid, stable, and transparent models that serve educational stakeholders effectively. This evolution is not merely academically desirable but increasingly legally necessary, as emerging regulatory frameworks impose mandatory transparency requirements for educational AI systems.

Future research should prioritize standardized protocols for evaluating sequential stability, systematic methods for addressing imbalance, and inherently interpretable architectures that embed transparency into core design. Expanding validation beyond common datasets to diverse educational contexts is essential for ensuring global applicability. DKT research should also integrate insights from cognitive science and learning theory to ensure that advances in modeling translate into meaningful educational outcomes.