betaFeedback

WorksComputer Science

Frac-BKT Under Mechanism Scrutiny: Response Assimilation and a Failed Fractional-Memory Test

Abstract

Long-horizon knowledge tracing requires more than an evocative memory mechanism: the implementation must assimilate observed responses non-anticipatorily, and the mechanism must survive matched controls. We subject Frac-BKT, a BKT-style model with an elapsed-time-scaled truncated fractional history, to both tests. The repaired model predicts response nn from a prior mastery state and uses response nn only to update the posterior state for later events. A five-seed evaluation uses 676,848 ASSISTments 2017 interactions from 1,694 learners and 19,249 learner–skill trajectories, with chronological 80/10/10 suffix splits. The response-aware model reaches AUC .5860 overall, .5638 after gaps of at least 30 minutes, and .5668 after at least 14 days. Relative to the earlier non-response implementation, the gains are .0473, .0316, and .0277 AUC, respectively. However, the fractional mechanism gate fails: the full model is not better than the matched α=1α=1 control and is worse than the matched τ=0τ=0 control by .0090 AUC on both long-gap slices. Student-cluster wild-bootstrap intervals exclude zero, both Holm-adjusted pp-values are .00006, and all five per-seed differences favor τ=0τ=0. Compact AKT-, SAKT-, and DAS3H-style references also remain stronger on long gaps. The contribution is therefore a response-assimilation repair and a falsifiable audit showing that response assimilation helps while the evaluated fractional-history term does not—not a state-of-the-art or causal-retention claim.

CCS Concepts • Applied computing → Interactive learning environments; • Computing methodologies → Supervised learning; • Mathematics of computing → Time series analysis.

Introduction

An educational model can be accurate on the next click and still be uninformative about what happens after the next week. This distinction is increasingly consequential as AI-mediated tutoring makes immediate assistance cheap and ubiquitous. A learner may receive productive scaffolding, or may obtain an answer without the retrieval and error diagnosis needed for durable learning. Immediate correctness alone cannot distinguish these cases. We call the resulting measurement risk—proximal success without verified durable competence—the mastery-gain problem. The phrase names an application motivation, not an empirical finding of this paper: ASSISTments 2017 contains legacy tutor hints rather than generative-AI interactions, and it does not contain randomized delayed post-tests.

Knowledge tracing (KT) is a natural component of such an audit because it separates a latent state from observed correctness. Bayesian Knowledge Tracing (BKT) is compact and interpretable, but its state transition is Markovian [13]. Deep sequence models learn richer histories and often predict better [12, 15, 30, 32], yet their internal states do not necessarily provide a small set of quantities that can be inspected as a specified learning mechanism. Time-aware KT methods address forgetting and spacing directly [11, 19, 29, 40], but there remains room for models that make the assumed memory law explicit.

Fractional dynamics offer one such inductive bias. A Caputo derivative represents the present rate of change as a power-law-weighted history of prior changes [7, 26, 33]. Unlike a one-step Markov transition, the resulting recurrence can distinguish trajectories with the same current state but different recent state increments. Unlike unconstrained attention weights, the history weights follow a declared analytic form. This does not make a fitted parameter a biological measurement, but it does make the model’s memory assumption inspectable.

We develop and stress-test an auditable fractional-state KT model, Frac-BKT. The claim-to-code audit materially narrows the claims relative to an accuracy-leaderboard paper. The canonical experiment predicts later suffix events from earlier events in the same learner–skill trajectory; it does not evaluate cold-start generalization to unseen learners. Its long-gap outcomes are next responses after elapsed-time gaps; they are retention-oriented stress tests rather than direct causal retention measurements. Its deep and time-aware comparators are compact, controlled in-pipeline references, not faithful reproductions of the cited systems.

Our contributions are:

  1. We repair the knowledge-tracing semantics by predicting each response from a prior state and applying an exact BKT emission posterior only after that prediction. A non-anticipation test verifies that changing response nn cannot change prediction nn but does change later predictions.

  1. We give the exact implemented solver: a truncated uniform-lag L1 history with an event-specific elapsed-time scale [37]. This makes clear both what the implementation computes and what it does not claim to approximate exactly.

  1. We report a reproducible five-seed chronological-suffix evaluation with null-centered student-cluster wild bootstraps and pre-declared next-session and 14+ day slices. The results include negative comparisons and calibration failures, not only favorable AUC differences.

(4) We enforce matched α=1\alpha= 1 and τ=0\tau= 0 mechanism gates under identical data, architecture, optimizer, seeds, and epochs. The completed controls falsify the fractional-history advantage: the no-history control is stronger on both pre-declared long-gap slices, with conditional learner-cluster intervals excluding zero.

Related Work

Knowledge Tracing and Time

BKT models skill mastery with a hidden state and separates mastery from correctness through guess and slip probabilities [5, 13, 34]. Its standard transition is memoryless after conditioning on the current state, a simplification that conflicts with evidence that learning depends on elapsed time, repetition, and spacing [10, 14, 27]. Deep Knowledge Tracing and transformer variants relax this restriction by learning sequence representations [12, 15, 30, 32, 43]. Benchmark work also shows that preprocessing, model-specified constraints, and implementation details can change model rankings, making protocol disclosure part of the scientific result [22, 25, 38, 42].

Time-aware KT is already a substantial field. DAS3H uses recency and spacing windows; HawkesKT models temporal excitation; forgetting-aware neural models inject elapsed-time signals or recency biases [11, 19, 29, 40]. Trainable spaced-repetition systems similarly connect elapsed time to review decisions [31, 35]. Frac-BKT is not offered as a replacement for this literature. Its narrower distinction is a latent-state transition with an explicit fractional history and a compact exported diagnostic state.

Assistance and Measurement

Feedback and mastery-learning research emphasize durable capability rather than task completion [8, 9, 16, 36]. Intelligent tutors operationalize that goal through diagnosis and scaffolding [4, 24, 39]. Help seeking is beneficial only under constraints: assistance can support learning, but overly direct help can enable shallow completion or gaming [2, 3, 6, 20]. Generative AI intensifies this measurement problem [18, 21]. However, legacy hint counts are not an identifying proxy for LLM use. We therefore use assistance telemetry only as a model input and descriptive diagnostic, not as a causal treatment.

Fractional Memory

For 0<α<10 < \alpha< 1, a Caputo derivative integrates prior rates of change with a power-law kernel; α=1\alpha= 1 recovers an ordinary first derivative [7, 26, 33]. Fractional operators have been used as long-memory priors in physical systems and differentiable learning [41]. Their main computational cost is history: direct evaluation is quadratic in sequence length. L1 discretization and finite-history truncation provide a practical recurrence [37], but irregular timestamps require care. In particular, scaling a uniform-lag L1 recurrence by the current gap is not the same as a mathematically derived nonuniform-grid L1 method [37]. We preserve that distinction throughout.

Model

Observed Trajectories

For learner uu, skill ss, and opportunity nn, the log contains correctness yu,s,n∈{0,1}y_{u,s,n} \in\{0,1\}, hint count hu,s,nh_{u,s,n}, attempt count au,s,na_{u,s,n}, and timestamp tu,s,nt_{u,s,n}. We model one trajectory per learner–skill pair. The latent state Mu,s,n∈[0,1]M_{u,s,n} \in[0,1] is a model-defined mastery state, not an independently observed test score.

We distinguish the prior state MnM_n, used to predict response nn, from the posterior state Mn+M_n^{+}, obtained only after observing that response. The implemented pedagogical dose shifts hints and attempts by one event:

Fpedn=1+ws(h)log⁡(1+hn−1)+ws(a)log⁡(1+an−1),(1)F_{\mathrm{ped}n} = 1 + w_s^{(h)} \log(1 + h_{n-1}) + w_s^{(a)} \log(1 + a_{n-1}), \tag*{(1)}

with zero counts at the first event and nonnegative weights obtained by softplus. A history context

gn=∑k<nhkmax⁡(∑k<nak,ϵ)(2)g_n = \frac{\sum_{k<n} h_k}{\max\left(\sum_{k<n} a_k,\epsilon\right)} \tag*{(2)}

summarizes cumulative assistance relative to attempts. We call this the prior-assistance ratio. The immutable canonical preprocessing artifact uses the legacy field name gaming_index; the repaired runner aliases that field before it enters the model. We infer no gaming behavior from this ratio.

Event-Conditioned Physics Parameters

A skill embedding ese_s and the prior context gng_n enter a two-layer network:

(α~n,κ~n,λ~n)=fθ([es,gn]).(3)(\tilde{\alpha}_n,\tilde{\kappa}_n,\tilde{\lambda}_n) = f_\theta([e_s,g_n]). \tag*{(3)}

The constrained outputs are

αn=clip⁡(σ(α~n),0.05,1),(4)\alpha_n = \operatorname{clip}(\sigma(\tilde{\alpha}_n), 0.05, 1), \tag*{(4)}
κn=softplus⁡(κ~n),λn=softplus⁡(λ~n).(5)\kappa_n = \operatorname{softplus}(\tilde{\kappa}_n), \qquad\lambda_n = \operatorname{softplus}(\tilde{\lambda}_n). \tag*{(5)}

Thus α\alpha, κ\kappa, and λ\lambda are event-conditioned outputs, not fixed skill constants. Reported slice values are averages over events and seeds. Within the specified model family, smaller αn\alpha_n gives a heavier-tailed lag kernel, κn\kappa_n scales conversion of dose into state gain, and λn\lambda_n applies decay pressure. These interpretations describe the recurrence; they do not establish cognitive constructs.

Elapsed-Time-Scaled Fractional Recurrence

The motivating continuous equation is

0CDtαM(t)=κFped(t)(1−M(t))−λM(t).(6){}^{C}_{0}D_t^{\alpha}M(t) = \kappa F_{\mathrm{ped}}(t)(1-M(t)) - \lambda M(t). \tag*{(6)}

For event nn, the implementation defines a transformed gap

δn=log⁡(1+max⁡(tn−tn−1,60 s)60 s)(7)\delta_n = \log\left(1 + \frac{\max(t_n-t_{n-1},60\ \mathrm{s})}{60\ \mathrm{s}}\right) \tag*{(7)}

and an event scale

Cn=δn−αnΓ(2−αn).(8)C_n = \frac{\delta_n^{-\alpha_n}}{\Gamma(2-\alpha_n)}. \tag*{(8)}

It retains uniform-lag L1 weights

bj(αn)=(j+1)1−αn−j1−αn(9)b_j(\alpha_n) = (j+1)^{1-\alpha_n} - j^{1-\alpha_n} \tag*{(9)}

for the most recent τ\tau posterior-mastery increments:

Hn,τ=∑j=1min⁡(n−1,τ)bj(αn)(Mn−j+−Mn−j−1+).(10)H_{n,\tau} = \sum_{j=1}^{\min(n-1,\tau)} b_j(\alpha_n)(M_{n-j}^{+} - M_{n-j-1}^{+}). \tag*{(10)}

The bounded prior-state recurrence is

Mn−=Π[0,1][CnMn−1+−CnHn,τ+κnFpednCn+κnFpedn+λn].(11)M_n^{-} = \Pi_{[0,1]}\left[\frac{C_nM_{n-1}^{+} - C_nH_{n,\tau} + \kappa_nF_{\mathrm{ped}n}}{C_n + \kappa_nF_{\mathrm{ped}n} + \lambda_n}\right]. \tag*{(11)}

At the first event, the unavailable prior-state and history terms are omitted.

Equation (11) is an elapsed-time-scaled, uniform-lag L1-inspired approximation [37]. It is not an exact nonuniform-grid discretization of Equation (6). The current gap changes CnC_n, while the history weights still depend on opportunity lag jj. We expose this hybrid explicitly because calling it continuous time would overstate the numerical method.

With fixed τ\tau, the recurrence costs O(Tτ)O(T\tau) per trajectory and stores O(T)O(T) states during training. Setting τ=0\tau= 0 removes Hn,τH_{n,\tau}. Setting αn=1\alpha_n = 1 annuls all bjb_j and yields the matched one-step architecture.

Non-Anticipatory Emission and Posterior

Skill-level guess and slip logits connect the prior state to the current response:

pn=p(yn=1)=M‾n(1−slips)+(1−M‾n)guesss.(12)p_n = p(y_n = 1) = \overline{M}_n(1-\mathrm{slip}_s) + (1-\overline{M}_n)\mathrm{guess}_s . \tag*{(12)}

After scoring pnp_n, the observed response updates the posterior by the exact BKT emission rule:

Mn+={M‾n(1−slips)pn,yn=1,M‾nslips1−pn,yn=0.(13)M_n^+ = \begin{cases} \dfrac{\overline{M}_n(1-\mathrm{slip}_s)}{p_n}, & y_n = 1, \\ \dfrac{\overline{M}_n\mathrm{slip}_s}{1-p_n}, & y_n = 0. \end{cases} \tag*{(13)}

Thus yny_n can affect only events n+1n+1 and later. An executable self-test changes one response and verifies equality of all predictions through that event and a change in later predictions. The run minimizes binary cross-entropy on training-suffix events:

L=−∑n∈Ωtrain[ynlog⁡pn+(1−yn)log⁡(1−pn)].(14)\mathcal{L} = -\sum_{n \in\Omega_{\mathrm{train}}} \left[y_n \log p_n + (1-y_n)\log(1-p_n)\right]. \tag*{(14)}

No additional regularizer is active. All models use final-epoch weights; there is no validation-selected checkpoint. This design is simple and reproducible, but it also limits claims about optimal model selection.

Evaluation

Data and Split

We use ASSISTments 2017 interaction logs [17, 23]. The raw file contains 942,816 rows; its SHA-256 digest is recorded in the artifact manifest. After retaining valid binary-response rows and learner–skill trajectories with at least 15 opportunities, 676,848 interactions remain: 1,694 learners, 81 skills, and 19,249 trajectories.

Each trajectory is sorted by timestamp and stable row identifiers, then split chronologically: the first 80% of events are training, the next 10% validation, and the final 10% test. This yields 533,756/67,155/75,937 rows. The first event in each trajectory is not scored. Learners and skills therefore occur across partitions; the generalization target is online prediction of later events within an existing learner–skill trajectory, not prediction for held-out learners.

Two pre-declared test slices operationalize long-horizon stress tests. The next-session slice contains 6,386 responses after gaps of at least 30 minutes (1,312 learners, 5,207 trajectories). The 14d+ day slice contains 4,683 responses (1,216 learners, 4,073 trajectories). These are delayed next-response outcomes. They are not randomized delayed post-tests.

Models and Training

The response-aware suite uses fixed configurations, seeds 42, 101, 2026, 8888, and 2145, 15 epochs, batch size 64, Adam, and final-epoch evaluation. The full model uses embedding dimension 16, learning rate 0.005, and τ=10\tau= 10.

The controlled references include a differentiable BKT implementation and a separate time-aware BKT whose transition depends on the transformed elapsed gap. Compact TransformerEncoder implementations provide SAKT-style and AKT-style references; a student/skill embedding model with five spacing windows provides a DAS3H-style reference; and a compact recurrent temporal-excitation model provides a HawkesKT-style reference. These names describe architectural inspiration. They are not official implementations of the cited systems, and their scores must not be read as published benchmark rankings.

Two publication-critical controls use exactly the response-aware architecture: Response-α1\alpha1 fixes αn=1\alpha_n = 1 with τ=10\tau= 10, and Response-τ0\tau0 retains dynamic αn\alpha_n but removes the history tail. They share response assimilation, preprocessing, embeddings, optimizer, learning rate, batch size, epochs, seeds, and aggregation with the full model. The earlier non-response Frac-BKT is retained only as a forensic baseline: it exposes the benefit of repairing response assimilation, not a matched test of fractional memory.

Metrics and Inference

We compute AUC, negative log-likelihood, Brier score, and 10-bin expected calibration error [28]. The main table reports the AUC of predictions averaged across five seeds for each event, matching the unit used for paired inference. For each of the six pre-specified full-versus-control comparisons, we form paired DeLong AUC influence values, sum their differences within learner, apply an HC1 finite-cluster scale, and draw 99,999 null-centered Rademacher wild-bootstrap replicates. Two-sided Monte Carlo pp-values use (e+1)/(B+1)(e+1)/(B+1), where ee counts draws at least as extreme as the observed studentized difference; step-down Holm adjustment covers all six comparisons. Inverted null-error quantiles give the reported 95% intervals. This inference remains conditional on the fitted seed-averaged ensemble. We therefore also export all five paired per-seed differences descriptively, without treating them as a population-level retraining test.

Results

Predictive Boundary

Table 1 separates model repair from ranking. Adding the ordered response update raises the earlier Frac-BKT by .0473 overall AUC (95% CI [.0434, .0512]), .0316 on 30m+ ([.0249, .0383]), and .0277 on 14d+ ([.0202, .0347]). It also exceeds differentiable BKT by .0593 overall, .0437 on 30m+, and .0408 on 14d+, with all intervals excluding zero. These are large effects of the repaired model as a whole; they do not isolate its fractional term.

ModelOverall AUC30m+ AUC14d+ AUC
AKT-style ref..6275.6313.6308
SAKT-style ref..5858.6126.6159
DAS3H-style ref..5925.6113.6062
HawkesKT-style ref..5901.5035.5081
Differentiable BKT ref..5267.5201.5260
Time-aware BKT ref..5242.5162.5220
Earlier non-response Frac-BKT.5387.5322.5391
Response-α1\alpha1.5855.5654.5695
Response-τ0\tau0.5872.5728.5758
Full response-aware Frac-BKT.5860.5638.5668

Table 1. ASSISTments test AUC after averaging five predictions per event. The 30m+ and 14d+ columns are next responses after at least the stated gap. Reference rows are compact in-pipeline implementations, not official benchmark reproductions.

The response-aware full model is not the strongest predictor. AKT-style exceeds it by .0415 overall (.0414 prior to four-decimal rounding) and .0640 on 14d+. SAKT-style and DAS3H-style are higher on both long-gap slices. The full model exceeds the compact HawkesKT-style reference after long gaps but not overall. These controlled results rule out a state-of-the-art ranking claim.

Mechanism Gate

The mechanism gate fails. Relative to α=1\alpha= 1, the full model differs by -.0016 AUC on 30m+ (95% CI [-.0042, .0010], Holm p=.44462p = .44462) and -.0027 on 14d+ ([-.0057, .0002], Holm p=.21960p = .21960); both intervals include zero and the per-seed signs are mixed. Relative to τ=0\tau= 0, it is worse by .0090 on 30m+ ([-.0123, -.0057], Holm p=.00006p = .00006) and .0090 on 14d+ ([-.0126, -.0054], Holm p=.00006p = .00006). All five paired seed differences favor τ=0\tau= 0 on each delayed slice. Thus learned α\alpha is not shown to help, and the evaluated history tail reduces long-gap discrimination under this protocol. The adjusted values quantify learner-cluster uncertainty conditional on the fitted five-seed ensemble, not uncertainty under retraining or alternative datasets.

Visualizing the Falsification

Figure 1 makes the publication gate explicit. The three response-aware models are close overall, but the no-history control separates on delayed slices. This conditional result uses predictions averaged over five seeds, the same non-anticipatory posterior update and training protocol, and a paired student-cluster wild bootstrap. Event-level α\alpha, κ\kappa, and λ\lambda remain available as internal diagnostics, but parameter variation cannot be interpreted as evidence for a mechanism that loses its matched ablation.

Figure 1. Matched response-aware evidence. (a) AUC of seed-averaged predictions. (b) Full-minus-control differences with 95% student-cluster wild-bootstrap intervals. The history-free τ=0\tau= 0 control is stronger on both long-gap slices, and both intervals exclude zero.

Discussion

What the Evidence Supports

The evaluation supports three claims. First, non-anticipatory response assimilation is essential: the repaired model substantially outperforms the earlier implementation that did not feed observed responses into future mastery. Second, the response-aware architecture beats both BKT-family references on all three reported slices. Third, the matched interventions identify a boundary that an uncontrolled comparison cannot: the tested fractional-history tail does not improve long-horizon discrimination.

The evaluation does not support three stronger claims. It does not establish state-of-the-art KT accuracy; compact attention and spacing references are stronger on long gaps. It does not directly measure retained knowledge; the outcome is correctness on the next logged response after a gap. It does not demonstrate that AI or hints cause mastery inflation; no generative-AI treatment or randomized delayed post-test is present.

This negative mechanism result is informative. A named parameter is not an interpretation, and a continuous-time motivation does not show that its discrete implementation helps. Mechanism-level interpretability requires matched interventions that can falsify the proposed explanation.

Implications for AI-Assisted Learning

The mastery-gain problem remains a useful deployment motivation. A tutoring system that optimizes only immediate response success can reward answer delivery rather than learning. A retention-oriented KT layer could help schedule reassessment or identify uncertainty after long gaps. However, the current model is not ready to generate learner risk flags. Its calibration is weak, its parameters are conditional model outputs, and its evidence comes from one legacy tutoring dataset. A responsible next study should combine randomized assistance policies, delayed post-tests, direct AI interaction logs, and pre-registered construct validation.

Limitations and Threats to Validity

Construct validity. Next-response-after-gap is only a proxy for retention. Latent MnM_n, αn\alpha_n, κn\kappa_n, and λn\lambda_n are internal to the model. Their labels do not make them independent cognitive measurements.

Internal validity. Hint and attempt counts are observational and may reflect prior mastery, item difficulty, curriculum structure, or help-seeking policy. We make no causal hint claim. The within-trajectory split conditions on earlier events from the same learner and should not be confused with learner-held-out generalization.

Numerical validity. The hybrid recurrence scales uniform-lag L1 weights with the current transformed gap. It is neither a constant-step experiment nor a derived nonuniform L1 solver [37]. A future numerical study should compare against a proper nonuniform-grid method and test the one-minute gap floor.

External validity. Results use one dataset, only trajectories with at least 15 opportunities, and 71.8% of raw rows. The in-pipeline reference models are not official benchmark implementations. A second dataset and a library-controlled benchmark are required before broad ranking claims.

Selection and uncertainty. Fixed configurations and final-epoch weights avoid post-hoc test selection (rather than automated hyperparameter search [1]) but do not provide equal validation-search budgets. The paired wild bootstrap operates on learner-cluster influence values after seed averaging. Per-seed effects expose observed optimization sensitivity, but neither analysis captures uncertainty from retraining on new samples, choosing another dataset, or changing model selection. be executable, temporally ordered, and subjected to matched interventions. An interpretable parameter earns scientific meaning only when the mechanism it names survives that audit.

Reproducibility and Ethics

The internal audit archive records the raw-file hash, preprocessing and split manifest, prediction and diagnostic files for all 15 response-aware model–seed jobs, 15 checkpoints, 225 unique history rows, metrics, and paired bootstrap outputs. The corrected inference archive adds all 30 per-seed comparisons, six raw and Holm-adjusted tests, 8,444 privacy-preserving cluster-influence rows, and 599,994 retained null draws. Completed jobs are written atomically and the runner is resumable. The non-anticipation self-test and exact α=1\alpha= 1 and τ=0\tau= 0 tests run before training. Dataset, canonical-runner, and response-runner hashes are stored in the completed metadata and independently rechecked.

The study analyzes de-identified educational interaction logs and does not intervene on learners. Its main ethical risk is inferential: presenting an internal model parameter as a learner trait could lead to inappropriate support or surveillance. We therefore avoid learner-facing labels, causal language about hints, and deployment claims. Any future use with generative-AI tutors should undergo institutional review, validate calibration and subgroup performance, and keep pedagogical decisions contestable by learners and educators.

Conclusion

The strongest result is not the one suggested by the model’s name. Non-anticipatory response assimilation produces a large and reproducible improvement over the earlier Frac-BKT implementation, but the full fractional-history model does not beat α=1\alpha= 1 and is worse than τ=0\tau= 0 after both 30 minutes and 14 days, with conditional learner-cluster intervals excluding zero. Attention and spacing references also remain stronger on long gaps. The broader lesson is methodological: knowledge-tracing mechanisms should

Acknowledgments

The author thanks Prof. Ricky Lay Kee Ang, Dr. Kee Chun-Yun, and Nguyen Gia Tuyen at the Singapore University of Technology and Design for motivating discussions on fractional dynamics and independent replication checks following the July 2026 canonical evaluation archived here.

Idealized L1 Derivation

For a uniform grid tn=nΔtt_n = n\Delta t, the Caputo derivative is

0CDtαM(tn)=1Γ(1−α)∫0tn(tn−ξ)−αM′(ξ) dξ.(15){}_{0}^{C}D_t^\alpha M(t_n) = \frac{1}{\Gamma(1-\alpha)}\int_{0}^{t_n}(t_n-\xi)^{-\alpha}M'(\xi)\,\mathrm{d}\xi. \tag*{(15)}

Approximating M′(ξ)M'(\xi) as constant on each interval gives the standard L1 discretization [37]:

0CDtαM(tn)≈Δt−αΓ(2−α)[(Mn−Mn−1)+∑j=1n−1bj(α)(Mn−j−Mn−j−1)].(16){}_{0}^{C}D_t^\alpha M(t_n) \approx\frac{\Delta t^{-\alpha}}{\Gamma(2-\alpha)}\left[(M_n-M_{n-1})+\sum_{j=1}^{n-1}b_j(\alpha)(M_{n-j}-M_{n-j-1})\right]. \tag*{(16)}

where bj(α)=(j+1)1−α−j1−αb_j(\alpha)=(j+1)^{1-\alpha}-j^{1-\alpha}. Substitution into Equation (6) and collection of terms in MnM_n yields the idealized implicit update corresponding to Equation (11). The canonical implementation replaces the common Δt\Delta t scale by the event-specific transformed gap in Equation (8) while retaining the uniform-lag weights. This last replacement is an engineering approximation, not part of the uniform-grid derivation.

Response-Aware Protocol Details

The artifact manifest records the 942,816-row raw file and its SHA-256 digest. The retained sample has 676,848 rows, 1,694 learners, 81 skills, and 19,249 trajectories. The split is chronological 80/10/10 within trajectory and excludes each trajectory’s first event from scoring. Training uses seeds 42, 101, 2026, 8888, and 12345 for 15 epochs with batch size 64 and final-epoch evaluation. The response-aware models use embedding dimension 16 and Adam with learning rate .005; the full and α=1\alpha=1 models use τ=10\tau=10, while the no-history control uses τ=0\tau=0. The corrected paired inference uses 99,999 null-centered Rademacher wild-bootstrap draws per comparison after seed averaging, with learner-clustered DeLong influence values, finite-Monte-Carlo correction, and step-down Holm adjustment. Note that while Table 1 reports the AUC of the seed-averaged ensemble predictions (.5860 for full response-aware Frac-BKT, .5855 for α=1\alpha=1, and .6275 for AKT-style), the corresponding arithmetic means of the five individual per-seed AUCs from the same July 2026 evaluation runs are .5853 for full response-aware Frac-BKT, .5845 for the matched α=1\alpha=1 control, and .6210 for the AKT-style reference. All 15 prediction files, 15 diagnostic files, 15 checkpoints, 225 history rows, the split manifest, metrics, bootstrap results, metadata, and hashes are retained in the internal audit archive. In the artifact glossary, long-horizon denotes a next response after a pre-declared gap, and named parameters denote event-conditioned recurrence outputs. Style references are compact in-pipeline architectures; matched controls change only α=1\alpha=1 or τ=0\tau=0 inside the same non-anticipatory response-aware implementation. None of these terms denotes a causal effect of hints or generative AI.

ControlSliceΔ\DeltaAUC [95% CI]Raw ppHolm ppSeeds
α=1\alpha=1Overall.00055 [-.00060, .00169].34486.444624/5
α=1\alpha=130m+-.00163 [-.00425, .00099].22231.444623/5
α=1\alpha=114d+-.00275 [-.00571, .00019].06820.219604/5
τ=0\tau=0Overall-.00114 [-.00230, .00002].05490.219602/5
τ=0\tau=030m+-.00902 [-.01230, -.00572].00001.000065/5
τ=0\tau=014d+-.00898 [-.01260, -.00538].00001.000065/5

Table 2. Corrected matched-control inference. Differences are full response-aware Frac-BKT minus control; Holm adjustment covers all six rows. Seed counts are descriptive.

References

  1. [1]Akiba, Takuya, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. 2019. Optuna: A Next-Generation Hyperparameter Optimization Framework. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. ACM, New York, NY, USA, 2623–2631. doi:10.1145/3292500.3330701
  2. [2]Vincent Aleven, Ido Roll, Bruce M. McLaren, and Kenneth R. Koedinger. 2016. Help Helps, but Only So Much: Research on Help Seeking with Intelligent Tutoring Systems. International Journal of Artificial Intelligence in Education 26, 1 (2016), 205–223.DOI
  3. [3]Vincent Aleven, Elmar Stahl, Silke Schworm, Frank Fischer, and Raven Wallace. 2003. Help Seeking and Help Design in Interactive Learning Environments. Review of Educational Research 73, 3 (2003), 277–320.DOI
  4. [4]John R. Anderson, Albert T. Corbett, Kenneth R. Koedinger, and Ray Pelletier. 1995. Cognitive Tutors: Lessons Learned. The Journal of the Learning Sciences 4, 2 (1995), 167–207.
  5. [5]Anirudhan Badrinath, Frederic Wang, and Zachary A. Pardos. 2021. pyBKT: An Accessible Python Library of Bayesian Knowledge Tracing Models. In Proceedings of the 14th International Conference on Educational Data Mining. International Educational Data Mining Society, Online, 468–474.arxiv.org/abs/2105.00385
  6. [6]Ryan S. Baker, Albert T. Corbett, and Kenneth R. Koedinger. 2004. Detecting Student Misuse of Intelligent Tutoring Systems. In Intelligent Tutoring Systems: 7th International Conference, ITS 2004 (Lecture Notes in Computer Science, Vol. 3220). Springer, Berlin, Heidelberg, 531–540. doi:10.1007/978-3-540-30139-4_50DOI
  7. [7]Dumitru Baleanu, Kai Diethelm, Enrico Scalas, and Juan J. Trujillo. 2016. Fractional Calculus: Models and Numerical Methods (2 ed.). Series on Complexity, Nonlinearity and Chaos, Vol. 5. World Scientific, Singapore. doi:10.1142/10044
  8. [8]Paul Black and Dylan Wiliam. 1998. Assessment and Classroom Learning. Assessment in Education: Principles, Policy & Practice 5, 1 (1998), 7–74.DOI
  9. [9]Benjamin S. Bloom. 1968. Learning for Mastery. Evaluation Comment 1, 2 (1968), 1–12.
  10. [10]Nicholas J. Cepeda, Harold Pashler, Edward Vul, John T. Wixted, and Doug Rohrer. 2006. Distributed Practice in Verbal Recall Tasks: A Review and Quantitative Synthesis. Psychological Bulletin 132, 3 (2006), 354–380. doi:10.1037/0033-2909.132.3.354
  11. [11]Benoît Choffin, Fabrice Popineau, Yolaine Bourda, and Jill-Jênn Vie. 2019. DAS3H: Modeling Student Learning and Forgetting for Optimally Scheduling Distributed Practice of Skills. In Proceedings of the 12th International Conference on Educational Data Mining. International Educational Data Mining Society, Montréal, QC, Canada, 29–38.arxiv.org/abs/1905.06873
  12. [12]Youngduck Choi, Youngnam Lee, Junghyun Cho, Jineon Baek, Byungsoo Kim, Youn-joong Cha, Dongmin Shin, Chanyoung Bae, and Jaewe Heo. 2020. Towards an Appropriate Query, Key, and Value Computation for Knowledge Tracing. In Proceedings of the Seventh ACM Conference on Learning @ Scale. ACM, New York, NY, USA, 341–344. doi:10.1145/3386527.3405945
  13. [13]Albert T. Corbett and John R. Anderson. 1994. Knowledge Tracing: Modeling the Acquisition of Procedural Knowledge. User Modeling and User-Adapted Interaction 4, 4 (1994), 253–278.DOI
  14. [14]Hermann Ebbinghaus. 1913. Memory: A Contribution to Experimental Psychology. Teachers College, Columbia University, New York, NY. Translated by Henry A. Ruger and Clara E. Bussenius; original work published 1885.
  15. [15]Aritra Ghosh, Neil T. Heffernan, and Andrew S. Lan. 2020. Context-Aware Attentive Knowledge Tracing. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. ACM, New York, NY, USA, 2330–2339. doi:10.1145/3394486.3403282
  16. [16]John Hattie and Helen Timperley. 2007. The Power of Feedback. Review of Educational Research 77, 1 (2007), 81–112.
  17. [17]Neil T. Heffernan and Cristina Lindquist Heffernan. 2014. The ASSISTments Ecosystem: Building a Platform that Brings Scientists and Teachers Together for Minimally Invasive Research on Human Learning and Teaching. International journal of Artificial Intelligence in Education 24, 4 (2014), 470–497.DOI
  18. [18]Wayne Holmes, Maya Bialik, and Charles Fadel. 2019. Artificial Intelligence in Education: Promises and Implications for Teaching and Learning. Center for Curriculum Redesign, Boston, MA.
  19. [19]Yoonjin Jin, Heejin Kook, Eunseong Choi, and Jongwuk Lee. 2023. Forgetting-Aware Linear Bias for Attentive Knowledge Tracing. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management. ACM, New York, NY, USA, 3958–3962. doi:10.1145/3583780.3615191DOI
  20. [20]Paul Salvador Inventado, Roberto O. Legaspi, and Jesus G. Bot icario. 2018. A Review of Research on Student Help Seeking and Help Design in Interactive Learning Environments. Artificial Intelligence Review 51, 2 (2018), 199–229.
  21. [21]Enkelejda Kasneci, Kathrin Sessler, Stefan Küchemann, Maria Bannert, Yaryna Dementieva, Frank Fischer, Urs Gasser, Georg Groh, Stephan Günnemann, Eyke Hüllermeier, Stephan Krusche, Gitta Kutyniok, Tilman Michaeli, Claudia Nerdel, Jürgen Pfeffer, Oleksandra Pqu et, Michael Sailer, Albrecht Schmidt, Tina Seidel, Matthias Stadler, Jochen Weller, Jochen Kuhn, and Gjergji Kasneci. 2023. ChatGPT for Good? On Opportunities and Challenges of Large Language Models for Education. Learning and Individual Differences 103 (2023), 102274.DOI
  22. [22]Mohammad M. Khajah, Robert M. Wing, Robert V. Lindsey, and Michael C. Mozer. 2016. How Deep is Knowledge Tracing? In Proceedings of the 9th International Conference on Educational Data Mining. International Educational Data Mining Society, Raleigh, NC, USA, 94–101.arxiv.org/abs/1604.02416
  23. [23]Kenneth R. Koedinger, Ryan S. J. Baker, Kyle Cunningham, Alida Skoghsholm, Brett Leber, and John Stamper. 2013. A Data Repository for the EDM Community: The PSLC DataShop. In Handbook of Educational Data Mining. CRC Press, Boca Raton, FL, USA, 43–56.
  24. [24]James A. Kulik and J. D. Fletcher. 2016. Effectiveness of Intelligent Tutoring Systems: A Meta-Analytic Review. Review of Educational Research 86, 1 (2016), 42–78.
  25. [25]Zitao Liu, Qiongqiong Liu, Jiahao Chen, Shuyan Huang, Jiliang Tang, and Weiqi Luo. 2022. pyKT: A Python Library to Benchmark Deep Learning Based Knowledge Tracing Models. In Advances in Neural Information Processing Systems, Vol. 35. Curran Associates, Inc., Red Hook, NY, USA, 18542–18555.arxiv.org/abs/2206.11460
  26. [26]Ralf Metzler and Joseph Klafter. 2000. The Random Walk’s Guide to Anomalous Diffusion: A Fractional Dynamics Approach. Physics Reports 339, 1 (2000), 1–77.DOI
  27. [27]Jaap M. J. Murre and Joeri Dros. 2015. Replication and Analysis of Ebbinghaus’ Forgetting Curve. PLOS ONE 10, 7 (2015), e0120644. doi:10.1371/journal.pone.0120644DOI
  28. [28]Mahdi Pakdaman Naeini, Gregory F. Cooper, and Milos Hauskrecht. 2015. Obtaining Well Calibrated Probabilities Using Bayesian Binning. In Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence. AAAI Press, Austin, TX, USA, 2901–2907.pubmed.ncbi.nlm.nih.gov/25927013
  29. [29]Koki Nagatani, Qian Zhang, Masahiro Sato, Yan-Ying Chen, Francine Chen, and Tomoko Ohkuma. 2019. Augmenting Knowledge Tracing by Considering Forgetting Behavior. In Proceedings of The World Wide Web Conference. ACM, New York, NY, USA, 3101–3107. doi:10.1145/3308558.3313565DOI
  30. [30]Shalini Pandey and George Karypis. 2019. A Self-Attentive Model for Knowledge Tracing. In Proceedings of the 12th International Conference on Educational Data Mining. International Educational Data Mining Society, Montréal, QC, Canada, 384–389.arxiv.org/abs/1907.06837
  31. [31]Harold Pashler, Greta Zarow, and Brian Triplett. 2003. Is Temporal Spacing of Tests Helpful Even When It Inflates Error Rates During Learning? Psychonomic Bulletin & Review 10, 4 (2003), 897–902. doi:10.3758/BF03196551
  32. [32]Chris Piech, Jonathan Bassen, Jonathan Huang, Surya Ganguli, Mehran Sahami, Leonidas J. Guibas, and Jascha Sohl-Dickstein. 2015. Deep Knowledge Tracing. In Advances in Neural Information Processing Systems, Vol. 28. Curran Associates, Inc., Red Hook, NY, USA, 505–513.arxiv.org/abs/1506.05908
  33. [33]Igor Podlubny. 1999. Fractional Differential Equations. Mathematics in Science and Engineering, Vol. 198. Academic Press, San Diego, CA.
  34. [34]Lawrence R. Rabiner. 1989. A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition. Proc. IEEE 77, 2 (1989), 257–286. doi:10.1109/5.18626DOI
  35. [35]Burr Settles and Brendan Meeder. 2016. A Trainable Spaced Repetition Model for Language Learning. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, Berlin, Germany, 1848–1858. doi:10.18653/v1/P16-1174DOI
  36. [36]Valerie J. Shute. 2008. Focus on Formative Feedback. Review of Educational Research 78, 1 (2008), 153–189.
  37. [37]Martin Stynes. 2022. A Survey of the L1 Scheme in the Discretisation of Time-Fractional Problems. Numerical Mathematics: Theory, Methods and Applications 15, 4 (2022), 1173–1192. doi:10.4208/nmtma. OA-2022-0009s
  38. [38]Eric G. Van Inwegen, Seth A. Adjei, Yutong Wang, and Joseph E. Beck. 2019. The Impact of Model-Specified Constraints in Knowledge Tracing. In Proceedings of the 12th International Conference on Educational Data Mining. International Educational Data Mining Society, Montréal, QC, Canada, 478–483.
  39. [39]Kurt VanLehn. 2011. The Relative Effectiveness of Human Tutoring, Intelligent Tutoring Systems, and Other Tutoring Systems. Educational Psychologist 46, 4 (2011), 197–221.DOI
  40. [40]Chenyang Wang, Weizhi Ma, Min Zhang, Chuancheng Lv, Fengyuan Wan, Huijie Lin, Taoran Tang, Yiqun Liu, and Shaoping Ma. 2021. Temporal Cross-Effects in Knowledge Tracing. In Proceedings of the 14th ACM International Conference on Web Search and Data Mining. ACM, New York, NY, USA, 517–525. doi:10.1145/3437963.3441811
  41. [41]J. Wang, Y. Wen, Y. Gou, Z. Ye, and H. Chen. 2017. Fractional-Order Gradient Descent Learning of BP Neural Networks with Caputo Derivative. Neural Networks 89 (2017), 19–30. doi:10.1016/j.neunet.2017.02.006
  42. [42]Xiaolu Xiong, Siyuan Zhao, Fei G. Van Inwegen, and Joseph E. Beck. 2016. Going Deeper with Deep Knowledge Tracing. In Proceedings of the 9th International Conference on Educational Data Mining. International Educational Data Mining Society, Raleigh, NC, USA, 545–550.
  43. [43]Yu Yin, Le Dai, Zhenya Huang, Shuanghong Shen, Fei Wang, Qi Liu, Enhong Chen, and Xin Li. 2023. Tracing Knowledge Instead of Patterns: Stable Knowledge Tracing with Diagnostic Transformer. In Proceedings of the ACM Web Conference 2023. ACM, New York, NY, USA, 855–864. doi:10.1145/3543507.3583255DOI

Paper details

vlnc:W_2x9jkjjs
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 8

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents