acceptodds
Under review as a conference paper at ICLR 2027

DSCA-KT: Dual-State Context-Aware Model for Dynamic Knowledge Tracing

Abstract

Knowledge Tracing (KT) aims to model students’ evolving knowledge states from historical learning interactions. Recently, dynamic KT methods have gained increasing attention for their ability to model continuously evolving interaction sequences. However, existing approaches still struggle to distinguish transient cognitive fluctuations from stable knowledge-acquisition patterns and often produce unreliable predictions in cold-start scenarios with limited interaction histories. To address these challenges, we propose DSCA-KT, a Dual-State Context-Aware framework for dynamic knowledge tracing. Specifically, we design context-aware recurrent aggregators that capture high-frequency cognitive variations, enabling the model to better characterize transient learning behaviors while preserving long-term knowledge trends. In addition, we introduce a context modulator that generates student- and question-specific contextual representations and incorporates prior structural knowledge to alleviate the cold-start issue. Extensive experiments on several widely used benchmark datasets demonstrate that DSCA-KT consistently outperforms state-of-the-art baselines under both transductive and inductive settings.

Primary area:
learning on time series and dynamical systems
Keywords:
Knowledge Tracing

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_eT1xOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper aims to address the challenges existing KT methods face in distinguishing short-term fluctuations from changes in students’ underlying knowledge states and handling cold-start scenarios with limited question's practice histories. The authors propose DSCA-KT, which jointly models student and question states using context-aware recurrent aggregators with frequency-based feature gating and a context modulator that incorporates response history, temporal information, and prior structure knowledge about questions and knowledge concepts.

Critical strengths

The paper addresses practically relevant challenges in knowledge tracing. The evaluation covers plenty of datasets under both transductive and inductive settings and includes extensive ablation studies to assess the contributions of the proposed components.

Critical weaknesses
  1. The connection between the proposed frequency-based operations and the claimed ability to distinguish meaningful cognitive changes remains insufficiently established. While the results support improved predictive performance, the evidence for this stronger mechanistic interpretation is less convincing.

  2. Although the paper evaluates inductive performance, it provides limited discussion of and comparison with KT methods specifically designed for cold-start scenarios.

Questions for authors
  1. Regarding the first weakness, Eq. 10–11 aggregates high-frequency magnitudes to generate a feature-wise gate g, which is then applied to X^Emb_s across all time steps. Could the authors clarify how this operation helps distinguish meaningful cognitive changes from transient noise? I appreciate that the method may work empirically, but further theoretical clarification would help support this interpretation.

  2. Regarding the second weakness, do any of the evaluated baselines specifically address cold-start scenarios? There are some existing works (e.g. csKT) explicitly studies cold-start KT. It would be better to include some comparisons with cold-start methods if the current experiments have not.

  3. Several confusions about Eq.2:

  • Eq. 2 appears to assign the indicators to the opposite sequences. In the student’s history, s_i = s always holds, while in the question’s history, q_i = q always holds. Should the student-side features use m_{q_i} and vice versa for the question-side?
  • In Eq. 2, the right-hand side uses a scalar indicator for a single step i, so the projected result x_ctx is d-dimensional. However, Eq. 3 adds this result to x_ans, whose dimension is N × d. The dimensions are mismatched here.
  1. Could the authors clarify the motivation for treating a question’s response history across different students as an order-sensitive sequence? While new responses provide additional evidence about a question, it is less clear what temporal changes the question-side state is intended to capture, or how high-frequency variations in this sequence should be interpreted. Such variations may reflect differences between the capabilities of students who practice, rather than changes in question state.

  2. As a minor presentation suggestion, please check the use of narrative and parenthetical citations throughout the paper. For example, “such as DKT Piech et al. (2015)” should read “such as DKT (Piech et al., 2015),” typically using \citep rather than \citet.

2 comments

Thank you very much for your further comments and advice. We will answer your questions and weaknesses point by point:

W1 & Q1: We clarify that FE uses high-frequency information as a conditioning signal for channel selection, rather than treating high-frequency content itself as evidence of meaningful cognitive change.

The spectral summary is computed from embeddings incorporating response-related and temporal context. Since the gate is shared across time steps, it does not perform event-level filtering. Instead, it adjusts each feature channel’s contribution based on the high-frequency magnitude statistics of the contextualized history. The reweighted embeddings retain their temporal ordering for subsequent GRU aggregation.

End-to-end predictive supervision learns these channel weights, providing the capacity to retain useful variation and attenuate less predictive features. Frequency alone cannot uniquely distinguish slips, insights, and noise; the gate does not independently identify such events.

Appendix A.5 illustrates responses to isolated errors, abrupt transitions, and alternating outcomes, rather than directly measuring cognitive processes. Our same-backbone experiments further show that FE exceeds Temporal-Std Gate by 1.12 and 1.13 AUC points on ASSIST2012 and ASSIST2017 under the transductive setting. This provides empirical evidence that spectral conditioning captures predictive information not fully reflected by the temporal-standard-deviation summary used in this control.

W2 & Q2: Thanks for the suggestion. csKT provides a relevant comparison specifically targeting cold-start KT. It uses kernel bias for generalization from short training sequences to longer test sequences and cone attention to capture hierarchical knowledge-component relationships. DSCA-KT differs in jointly modeling student- and question-side interaction histories while using a static knowledge graph to generate concept-derived priors for questions, particularly when their interaction histories are sparse.

We conducted an additional comparison on ASSIST2017 under a shared new-student evaluation protocol. Both models use the same data split and test interactions, including students’ first interactions. Predictions are grouped by the number of student interactions observed strictly before each prediction timestamp. This evaluates performance across student-history lengths rather than reproducing csKT’s original train-short/test-long protocol.

Student History LengthcsKT AUC (%)DSCA-KT AUC (%)csKT AP (%)DSCA-KT AP (%)
073.275.664.867.1
1–276.677.267.268.8
3–577.278.969.970.6
6–1079.080.171.072.1

DSCA-KT achieves higher AUC and AP than csKT across all student-history groups, with the largest gains in the zero-history group.

Q3: Sorry for the confusion. The indicators are relative to the target student–question pair. In the student’s history, the relevant indicator is mqi=I(qi=q)m_{q_i}=\mathbb{I}(q_i=q); in the question’s history, it is msi=I(si=s)m_{s_i}=\mathbb{I}(s_i=s). The implemented correspondence is therefore

\qquad x_{\mathrm{ctx},i}^{q}=m_{s_i}W_{\mathrm{st}}+b_{\mathrm{st}}.$$ These encode whether each historical interaction involves the target question or target student, respectively. The indicator subscripts in Eq. (2) are interchanged; the implementation already uses the correspondence above. Regarding dimensions, each projected $x_{\mathrm{ctx},i}$ is a $d$-dimensional row vector. Stacking the projections over the $N$ historical positions yields $$x_{\mathrm{ctx}}^{s},x_{\mathrm{ctx}}^{q}\in\mathbb{R}^{N\times d},$$ matching the answer embeddings in Eq. (3). Consequently, $CC_s$ and $CC_q$ are also $N\times d$ matrices before recurrent aggregation. These are notation and dimension-labeling issues; the implemented projections and additions operate consistently over the sequence dimension.

Q4: Sorry for the confusion. The question-side state represents the observed interaction context of a question up to the prediction time, rather than a cognitive state of the question itself. Its representation can evolve as evidence accumulates even when the question’s content and intrinsic properties remain unchanged.

Chronological aggregation captures recency and temporal patterns: histories with identical overall correctness rates can exhibit different recent response patterns. Chronological ordering, together with encoded time intervals, preserves each observation’s temporal proximity to the prediction time, allowing the model to distinguish recent patterns from older evidence. These features inform prediction jointly with the target student’s representation.

Variations across a question’s responses can indeed reflect differences in participating students’ abilities. Accordingly, question-side high-frequency statistics describe rapid variation in contextualized interaction features, without uniquely attributing it to intrinsic question difficulty. The frequency-conditioned gate adjusts channel contributions based on this variation; it does not identify its causal source. The motivation is to learn a temporally informed question representation complementary to student-side history, rather than assume that every response fluctuation reflects a change in the question itself.

Q5: Thanks for your suggestion. We will standardize citation formatting throughout the manuscript, using \citep for parenthetical references and \citet when author names form part of the sentence. The example should read “such as DKT (Piech et al., 2015).”

Reviewer_b7GuOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DSCA-KT, a dual-state context-aware framework for dynamic knowledge tracing. The model constructs learner-side and question-side historical interaction sequences and encodes response outcomes, repetition-related indicators, and temporal context. A frequency-domain module summarizes high-frequency magnitudes to generate a channel gate before recurrent aggregation with residual connections and layer normalization. The framework also incorporates structural priors from a static question–concept and concept-prerequisite graph to support question-side cold starts. Experiments compare DSCA-KT with twelve baselines on five public datasets under transductive and inductive settings, supplemented by ablations, spectral analyses, cold-start transfer experiments, and computational cost comparisons.

Critical strengths

1-Modeling transient response fluctuations and predicting performance on questions with limited interaction histories are worthwhile problems. Combining learner-side and question-side histories with structural priors is reasonably motivated. The experiments transferring the prior-context module to DTransformer, DyGFormer, and DyGKT in Figure 5 also provide encouraging evidence that this component may be useful beyond the proposed backbone.

2-Table 2 covers five public datasets and twelve baselines under both transductive and unseen-question settings. DSCA-KT achieves the highest reported mean AUC in all dataset–setting combinations. The paper reports variability and paired significance tests, and includes component ablations and efficiency comparisons. These results suggest potential practical value, subject to the methodological and evaluation-protocol concerns below.

Critical weaknesses

1-FlucKT already studies the separation and dynamic reweighting of long-term cognitive trends and short-term fluctuations through a causal-convolution-based decomposition layer. Although FlucKT is included in Table 2, Section 2 discusses it mainly as an attention-based model without engaging with this directly relevant contribution. The dynamic-graph setting and the proposed frequency-conditioned gate are different design choices, but the paper needs to explain more precisely what they add beyond existing fluctuation modeling.

More importantly, Eqs. (10)–(11) average high-frequency magnitudes into a channel-level vector and produce a gate shared across the entire input history. For a fixed input and channel cc, the operation therefore satisfies RFFT⁡(XSE)f,c=gcRFFT⁡(XEmb)f,c\operatorname{RFFT}(X^{\mathrm{SE}}){f,c}=g_c\operatorname{RFFT}(X^{\mathrm{Emb}}){f,c}: every frequency within that channel is scaled by the same factor. This is a channel-reweighting mechanism conditioned on high-frequency statistics, but the equations alone do not establish frequency-selective amplification within a channel. Reweighting channels with different spectral characteristics may still be useful; however, the connection to the frequency-localized filtering described around Figure 4 requires an explicit explanation.

Appendix A.4.2 does not resolve this gap. It infers ∂g/∂hˉ>0\partial g/\partial\bar h>0 from end-to-end BCE training, although no sign or monotonicity constraints are imposed on the projection weights. Optimizing prediction accuracy does not guarantee this derivative is positive. The synthetic examples in Appendix A.5 are informative illustrations, but they do not prove the stated monotonicity proposition or establish the claimed spectral compensation mechanism. These issues directly affect the interpretation of the paper’s central contribution.

2-Section 4.1 describes an 80/10/10 train/validation/test partition, while Section 4.2 reports results across five folds. The manuscript does not explain how these folds are constructed, what units are partitioned, or how temporal ordering and held-out-question selection are handled within each fold. This matters particularly because DSCA-KT uses both learner-side and question-side histories. I acknowledge that Section 3.2 restricts sampled neighbors to earlier timestamps and that unseen questions are excluded from SKG training. However, the split construction and the rules for incorporating newly observed interactions during validation and testing remain unclear.

The baseline tuning and model-selection procedures are also insufficiently documented. The paper provides several fixed training hyperparameters, but does not clarify whether these apply to all models or only DSCA-KT, nor does it report model-specific search spaces, tuning budgets, checkpoint-selection rules, or history-length settings. Appendix A.1 examines two hyperparameters on ASSIST17 and uses the resulting curves to choose their values, without identifying whether those curves are based on validation or test performance. Sensitivity analysis alone does not establish that the competing methods received a comparable validation-based tuning effort. These reporting gaps limit confidence that the observed gains can be attributed to the proposed mechanisms under a consistently controlled evaluation.

Questions for authors

1-Please explain how the sequence-shared gate in Eq. (11) produces the frequency-localized behavior described around Figure 4, including how the plotted quantities and reported gating values are computed. What assumptions or constraints justify the positive derivative in Eq. (21)? If monotonicity is not guaranteed, which theoretical conclusions remain valid?

2-Please discuss FlucKT’s decomposition and reweighting of long-term trends and short-term fluctuations directly, and identify the distinct contribution of DSCA-KT’s frequency-conditioned gate. Do the existing ablations distinguish the benefit of high-frequency conditioning from that of a comparable channel gate using non-spectral summaries? A focused comparison under the same backbone and input context would help clarify this point.

3-Please specify how the 80/10/10 partition and five folds are constructed, including temporal ordering, held-out-question sampling, and history-update rules during evaluation. Please also clarify how baselines handle unseen questions and report their implementation sources, tuning spaces and budgets, validation-based model-selection criteria, and history-length settings. Which data split was used to select the hyperparameters in Appendix A.1?

2 comments

Thank you very much for your further comments and advice. We will answer your questions and weaknesses point by point:

Q1: To explain how Figure 4 is computed, we randomly sample N=1,000N=1{,}000 histories from the ASSIST12 test set and extract the paired features:

XSE=XEmb⊙g,g∈RN×1×D.X^{SE}=X^{Emb}\odot g,\qquad g\in\mathbb{R}^{N\times1\times D}.

Since the gate is shared across sequence positions:

∣XSE^n,k,c∣=gn,c∣X^n,k,cEmb∣.|\widehat{X^{SE}}_{n,k,c}|=g_{n,c}|\widehat{X}^{Emb}_{n,k,c}|.

We average magnitudes across samples and channels:

Apre(k)=1ND∑n,c∣X^n,k,cEmb∣,Apost(k)=1ND∑n,cgn,c∣X^n,k,cEmb∣.A_{\mathrm{pre}}(k)=\frac{1}{ND}\sum_{n,c}|\widehat{X}^{Emb}_{n,k,c}|,\qquad A_{\mathrm{post}}(k)=\frac{1}{ND}\sum_{n,c}g_{n,c}|\widehat{X}^{Emb}_{n,k,c}|.

For visualization, the post-FE spectrum is DC-aligned:

A~post(k)=Apost(k)Apre(0)Apost(0).\widetilde{A}_{\mathrm{post}}(k)=A_{\mathrm{post}}(k)\frac{A_{\mathrm{pre}}(0)}{A_{\mathrm{post}}(0)}.

Displayed increases therefore reflect greater retention relative to DC, rather than amplification of unnormalized magnitudes. This effect arises from reweighting channels with different spectral profiles.

The reported gating statistics are computed before DC alignment as R(k)=Apost(k)/Apre(k)R(k)=A_{\mathrm{post}}(k)/A_{\mathrm{pre}}(k), where denominators are nonzero. This ratio is an input-magnitude-weighted average of channel gates, rather than a directly learned frequency-specific gate. We average R(k)R(k) over the selected peak neighborhoods and remaining high-frequency indices, respectively. These groups describe spectral regions, not independently annotated cognitive events.

Regarding the positive derivative in Eq. (21), including the implementation’s bias terms gives:

g=σ(W2ReLU⁡(W1hˉ+b1)+b2).g=\sigma(W_2\operatorname{ReLU}(W_1\bar{h}+b_1)+b_2).

Where ReLU is differentiable:

∂g∂hˉ=diag⁡(g⊙(1−g))W2diag⁡(I[W1hˉ+b1>0])W1.\frac{\partial g}{\partial\bar{h}}=\operatorname{diag}(g\odot(1-g))W_2\operatorname{diag}(\mathbb{I}[W_1\bar{h}+b_1 > 0])W_1.

Its entries can be positive, zero, or negative. Elementwise nonnegative weights would suffice for coordinatewise nondecreasing gates; locally strict positivity additionally requires active positive-weight paths. Our implementation imposes no such constraints, and BCE alone does not guarantee monotonicity. We will formulate Eq. (21) through this general Jacobian.

Without monotonicity, exact channel scaling, non-amplification of unnormalized FE-output magnitudes, and potentially frequency-dependent aggregate retention remain valid. Eq. (22) applies under a frozen-gate, locally linearized recurrent approximation; it omits the additional sensitivity from the gate’s input dependence. The recurrent low-pass interpretation likewise requires the corresponding state-transition assumptions. The supported mechanism is therefore high-frequency-conditioned channel modulation, whose predictive usefulness is supported by the controlled experiments. We will revise the monotonicity statement accordingly.

W1 & Q2: Thanks for the suggestion. FlucKT separates long-term trends and short-term fluctuations through causal convolution, recombines them using learned coefficients, and introduces a kernelized positional bias in attention. DSCA-KT instead summarizes high-frequency magnitudes from the current history and maps this summary through an MLP to generate channel weights before recurrent aggregation. FlucKT’s decomposition coefficients are learned model parameters, whereas DSCA-KT’s gate is generated for each input history according to its spectral content. This conditioning operates on both student and question interaction histories, allowing their representations to adapt to their respective temporal patterns within the dual-state framework.

To examine whether the benefit comes specifically from spectral conditioning, we conducted controlled experiments on ASSIST2012 and ASSIST2017 under the transductive setting. We kept the backbone and input context unchanged and compared removing FE with replacing its spectral summary by temporal mean or standard deviation, retaining the same gating architecture for the latter two variants.

VariantASSIST2012 AUC (%)ASSIST2012 AP (%)ASSIST2017 AUC (%)ASSIST2017 AP (%)
w/o FE80.15±1.0287.62±0.6179.31±0.0972.05±0.04
Temporal-Mean Gate80.25±1.0087.70±0.5979.84±0.0772.16±0.03
Temporal-Std Gate80.33±0.9987.78±0.5880.53±0.0672.25±0.03
DSCA-KT FE81.45±0.9888.89±0.5681.66±0.0573.34±0.02

FE consistently outperforms both non-spectral gates in mean AUC and AP, supporting the added value of high-frequency conditioning under the same backbone and input context.

W2 & Q3: Thanks for the suggestion. We clarify the data partitioning, evaluation-time histories, baseline configurations, and model-selection procedures below.

Data partition and repeated runs. The 80th and 90th timestamp percentiles define the boundaries between the training, validation, and test periods. The five results reported in Section 4.2 correspond to five independent training runs with different random seeds on a fixed chronological split, rather than five-fold cross-validation. For inductive evaluation, held-out question nodes are sampled using random seed 2020, with a sampling ratio of 10% of all unique question nodes. All interactions involving these held-out questions are excluded from training. Regular evaluation covers the full test window, whereas inductive evaluation focuses on test interactions involving a question unseen during training.

History construction and evaluation-time updates. For each target student and question, we retrieve at most 100 historical neighbors, ordered chronologically and restricted to timestamps strictly earlier than the target interaction. Interactions sharing the same timestamp are treated as simultaneous events and do not enter one another’s histories. During validation and testing, previously observed interactions can inform subsequent predictions, while current and future interactions are excluded. Model parameters remain fixed throughout evaluation.

Baseline implementations and training configurations. Each baseline follows its corresponding implementation and default model-specific configurations within the shared evaluation framework. The embedding dimension and dropout rate are set to 256 and 0.1, respectively, where applicable. Model-specific optimization and architectural hyperparameters follow their respective defaults. No additional baseline-specific hyperparameter search is conducted; the additional tuning budget is therefore zero.

ConfigurationSetting
Maximum historical neighbors per node100 (where applicable)
Embedding dimension / dropout256 / 0.1 (where applicable)
DSCA-KT optimizerAdam
DSCA-KT learning rate / weight decay5×10−45\times10^{-4} / 10−410^{-4}
DSCA-KT maximum epochs200
DSCA-KT validation intervalEvery 10 epochs
DSCA-KT early-stopping patience20 validation checks

Unseen-question handling. DSCA-KT generates prior representations for unseen questions by aggregating learned embeddings of their associated concepts through the static knowledge graph and CRE, then incorporates subsequently observed interaction histories. DyGKT uses projected raw node features and historical-neighbor aggregation, whereas DyGFormer uses its default node-feature and temporal-neighbor encoding. Other baselines follow their default initialization and update mechanisms without DSCA-KT’s prior module.

Model selection and Appendix A.1. Checkpoint selection uses validation AUC exclusively: the checkpoint with the highest validation AUC is selected for final testing, while AP is reported as a secondary evaluation metric. Test-set performance is not used for checkpoint or hyperparameter selection. Appendix A.1 uses the ASSIST17 validation set under the transductive setting, examining time thresholds of {1,12,24,48,72}\{1,12,24,48,72\} hours and high-frequency cutoffs of {0.4,0.5,0.6,0.7,0.8}\{0.4,0.5,0.6,0.7,0.8\}. Based on validation performance, we select 24 hours and 0.7, respectively.

Reviewer_aU93OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DSCA-KT, a dynamic knowledge tracing framework that models student and question states through context-aware recurrent aggregators and structural knowledge priors. The method achieves consistent AUC improvements across five datasets under both transductive and inductive settings. The experimental coverage is a strength, although several aspects of the technical justification require further attention.

Critical strengths

None.

Critical weaknesses
  1. Equations (10) and (11) aggregate high-frequency information into a channel-wise gate applied across the entire sequence. It is unclear how this operation selectively amplifies high-frequency components rather than simply rescaling the input features. Furthermore, the monotonicity claim in Appendix A.4 is not guaranteed by the unconstrained network weights. The authors should clarify the mathematical interpretation and reconcile it with the analysis in Figure 4.

  2. The inductive setting primarily considers unseen questions whose associated concepts remain available. Moreover, the prerequisite graphs are constructed through LLM-assisted annotation and human verification, which introduces additional information and preparation costs. The authors should examine whether the improvements persist when prerequisite relations are incomplete or inaccurate and clarify the contribution of this external information.

  3. The paper should discuss Time-frequency domain coupling-oriented knowledge tracing (TFKT, 2026), which also models learner states through frequency-domain decomposition. Other relevant studies include HiTSKT (2024), which considers learning dynamics across sessions, and BDGKT (2026), which models bidirectional student-question interactions. A clearer comparison would help establish the distinct contribution of DSCA-KT.

Questions for authors

Pls see the critical weakness section.

2 comments

Thank you very much for your further comments and advice. We will answer your questions and weaknesses point by point:

W1: Sorry for the confusion. FE performs high-frequency-conditioned channel reweighting, rather than independently manipulating individual frequency bins. For a fixed historical sequence X=XEmbX=X^{Emb}, the gate is shared across time: XSE^k,c=gc(X)X^k,c.\widehat{X^{SE}}_{k,c}=g_c(X)\widehat{X}_{k,c}. Each channel is therefore scaled uniformly across frequencies. Channels with different spectral profiles can nevertheless change their relative contributions to the aggregate spectrum. Since 0<gc<10 < g_c < 1, FE does not increase unnormalized spectral magnitudes at its output.

Figure 4 displays DC-aligned average magnitude spectra. Let Apre(k)A_{\mathrm{pre}}(k) and Apost(k)A_{\mathrm{post}}(k) denote magnitudes averaged across sampled histories and channels, and define R(k)=Apost(k)/Apre(k)R(k)=A_{\mathrm{post}}(k)/A_{\mathrm{pre}}(k). After aligning the post-FE spectrum to the pre-FE DC magnitude, A~post(k)Apre(k)=R(k)R(0),\frac{\widetilde{A}_{\mathrm{post}}(k)}{A_{\mathrm{pre}}(k)}=\frac{R(k)}{R(0)}, where the denominators are nonzero. Displayed increases thus indicate greater retention relative to DC, rather than absolute amplification. We have clarified the DC-aligned interpretation in Section 4.3.

Regarding Appendix A.4, the gate Jacobian depends on the learned weights and activation pattern; unconstrained weights and BCE supervision do not guarantee positive sensitivity. We will express Eq. (21) through the general Jacobian. The supported interpretation is input-dependent channel reweighting with potentially frequency-dependent aggregate retention, rather than guaranteed monotonic compensation of recurrent attenuation. Its predictive usefulness is supported by the controlled experiments.

W2: Thanks for the suggestion. Our cold-start objective concerns new questions associated with an existing concept vocabulary. The existing w/o PC ablation examines whether the framework’s performance depends entirely on the prior-context module. To further assess sensitivity to prerequisite quality, we conducted additional experiments on ASSIST2017 that randomly drop 20%/50% of prerequisite edges or perturb 10%/20% of them, while retaining question–concept associations. We also include a Q–C-only variant to isolate the contribution of prerequisite relations. Cold-Start Subset AUC is evaluated on the first 10 interactions of each question within the test split, while Overall AUC is evaluated on the full test split.

SettingCold-Start Subset AUC (%)Overall AUC (%)
DSCA-KT (Full Prior)77.13±0.1281.66±0.05
Random Drop 20% Prerequisites76.72±0.1581.02±0.08
Random Drop 50% Prerequisites76.16±0.3980.29±0.31
Perturb 10% Prerequisites76.54±0.1781.04±0.09
Perturb 20% Prerequisites75.87±0.3381.15±0.23
DSCA-KT (Q–C Only, No Prerequisites)75.69±0.1880.62±0.10
Strongest Baseline75.26±0.1680.35±0.14

Prerequisite information remains beneficial on the evaluated question subset under all tested corruptions, although substantial edge removal eliminates its overall advantage over Q–C Only and the baseline.

W3: Thanks for the suggestion. These studies are relevant, but differ from DSCA-KT in their modeling mechanisms.

TFKT uses a causal Fourier network with learnable masks to decompose knowledge states into multiple frequency bands for time–frequency coupling. DSCA-KT uses high-frequency summaries to condition channel reweighting before recurrent aggregation, applied to both student-side and question-side histories.

HiTSKT captures within-session and across-session dynamics through hierarchical Transformer encoders. DSCA-KT constructs timestamped histories around students and questions without requiring session segmentation, integrating temporal context with frequency-conditioned aggregation.

BDGKT models bidirectional student–question information flow through group-based dynamic graphs and distinct propagation mechanisms. DSCA-KT emphasizes context-conditioned aggregation of the two historical views, together with prior context for question-side cold starts.

To complement these architectural distinctions, we conducted additional comparisons on ASSIST2012 and ASSIST2017 under both transductive and inductive settings, reporting AUC and AP below. All values are percentages.

SettingModelASSIST2012 AUCASSIST2017 AUCASSIST2012 APASSIST2017 AP
TransductiveHiTSKT78.52±0.1578.14±0.1987.48±0.1470.42±0.18
TFKT79.35±0.1879.68±0.1687.96±0.1571.65±0.15
BDGKT79.62±0.2180.35±0.1488.15±0.1272.28±0.21
DSCA-KT81.45±0.9881.66±0.0588.89±0.5673.34±0.02
InductiveHiTSKT77.85±0.2477.42±0.2186.95±0.1869.85±0.24
TFKT78.82±0.2878.95±0.2287.58±0.1971.12±0.20
BDGKT79.45±0.2680.18±0.1888.05±0.1572.15±0.22
DSCA-KT81.07±0.8681.38±0.1189.02±0.7373.65±0.01

Dear Reviewer aU93,

As the discussion phase is approaching its end, we kindly request you to let us know if the above clarifications and the previously added experiments have addressed the remaining questions. If you are satisfied, we kindly request that you consider updating the score to reflect the newly added results and discussion. We would be happy to address any additional points the reviewer may have during the remaining time of the discussion phase. We thank the reviewer for engaging in the discussion with us.

Thank you for your time and consideration.

Best regards, The Authors

Reviewer_mJTEOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DSCA-KT, a Dual-State Context-Aware framework for dynamic knowledge tracing. Existing dynamic KT methods struggle to distinguish transient cognitive fluctuations (e.g., sudden insights or careless mistakes) from stable knowledge-acquisition patterns, and often produce unreliable predictions in cold-start scenarios with limited interaction histories. To address these challenges, the authors design (1) a context-aware recurrent aggregator that captures high-frequency cognitive variations via a frequency-domain enhancer, characterizing transient learning behaviors while preserving long-term knowledge trends, and (2) a context modulator that generates student- and question-specific contextual representations and incorporates structural knowledge priors from a static knowledge graph to alleviate the cold-start issue. Extensive experiments on several widely used benchmark datasets demonstrate that DSCA-KT consistently outperforms state-of-the-art baselines.

Critical strengths

1.This paper identifies that standard RNN-based aggregators in dynamic knowledge tracing act as low-pass filters that oversmooth transient cognitive fluctuations (e.g., sudden insights or careless mistakes). Modeling dynamic KT contexts in the frequency domain is a novel and well-motivated direction, supported by the spectral bias literature.

2.The evaluation is comprehensive and rigorous, covering 5 public datasets, 12 baselines spanning four methodological families, and both transductive and inductive settings with 5-fold reporting and statistical significance tests (paired t-test, p < 0.05). DSCA-KT achieves consistent AUC improvements (average +1.12%) over all state-of-the-art baselines.

3.The lifecycle-fading mechanism for the prior context is a thoughtful design: it honestly addresses the fact that static structural priors become redundant as dynamic representations accumulate, and gracefully transitions from prior-guided to interaction-driven modeling. Moreover, the plug-in experiment that integrates the prior context generator into three representative backbones demonstrates its generality as a cold-start solution without altering the core mechanisms of the hosts.

Critical weaknesses

1.There have been existing studies on disentangling transient cognitive fluctuations from stable knowledge states in knowledge tracing. The paper does not clarify the advantages of the proposed approach over these related works.

2.The novelty is limited. The frequency-domain enhancer is a new combination, but the other components are largely assembled from mature techniques: LayerNormGRU is simply GRU with residual connections and layer normalization, the prior context generator is a standard two-layer RGCN, and the temporal encoder uses a conventional bi-scale interval projection. The contribution reads more like a recombination of existing components than a fundamentally new design.

3.Modeling students' cognitive states in the frequency domain lacks pedagogical grounding. The paper does not provide any educational theory or learning-science evidence explaining why transient cognitive behaviors (e.g., insights or careless mistakes) should manifest as high-frequency signals, and why amplifying them is pedagogically meaningful.

4.The concept-concept prerequisite edges are constructed via LLM-assisted annotation with human verification. The reliability of this data is questionable: LLM-generated prerequisite relations may contain errors or curricula-specific biases, and the resulting graphs may not generalize across different curricula, which undermines the credibility of the prior context built upon them.

5.The average improvement of 1.12% in AUC is modest. On several datasets, the margin over the strongest baselines is only 1–2 points, which raises concerns about the practical significance of the gains.

6.The claimed cold-start solution appears to address only the question-side cold-start problem (new questions appearing). It remains unclear how the framework handles the concept-side cold-start problem (new concepts appearing) or the student-side cold-start problem (new students entering the system), both of which are common in real-world educational scenarios.

Questions for authors

1.Previous KT studies have attempted to separate transient cognitive behaviors from stable mastery states. Could the authors specify which of these works are most relevant to DSCA-KT, and clarify in what concrete aspects the proposed frequency-domain approach is superior?

2.The core assumption is that transient cognitive fluctuations correspond to high-frequency components. Is there any learning-science or psychometric theory supporting this mapping?

3.Since the prior context depends on the SKG, how sensitive is the model's performance to the quality of the prerequisite edges? For example, what happens if the concept-concept prerequisite edges are removed (keeping only question-concept inclusion edges), or if a fraction of edges is randomly perturbed?

4.The paper addresses question-side cold start. How does DSCA-KT behave when a new concept or a new student appears with no interaction history at all? Does the framework degrade gracefully in these cases, and is there any experimental evidence?

2 comments

Thank you for your comments. We address each concern below.

W1 & Q1: FlucKT and TFKT are closely related, but use frequency information differently from DSCA-KT. FlucKT separates trends and fluctuations through causal convolution, recombines them using learned coefficients, and introduces kernelized bias attention to strengthen local fluctuation modeling. TFKT applies causal FFT and learnable masks to decompose knowledge states into low-, middle-, and high-frequency components for time–frequency coupling.

DSCA-KT instead extracts a high-frequency magnitude summary from each current history and maps it to input-dependent channel weights. It modulates historical features before GRU aggregation, without reconstructing separate frequency-band states. This mechanism operates on both student and question interaction histories, allowing each view’s aggregation to adapt to its observed spectral characteristics.

We distinguish two forms of empirical evidence. Whole-model comparisons: Table 2 reports higher mean AUC than FlucKT across all ten dataset–setting combinations. Additional comparisons with TFKT on ASSIST2012 and ASSIST2017, under both transductive and inductive settings, yield mean AUC gains of 1.98–2.43 percentage points. These results demonstrate the predictive competitiveness of the complete framework, including its context modulator and structural priors.

Controlled mechanism comparison: On ASSIST2012 and ASSIST2017 under the transductive setting, keeping the backbone, input context, and gating architecture unchanged, FE exceeds Temporal-Std Gate by 1.12 and 1.13 AUC points, respectively. This comparison specifically supports the additional value of high-frequency conditioning over gating based on temporal variability.

W2: We agree that GRUs, residual connections, layer normalization, RGCNs, and bi-scale temporal projections are established techniques. We do not claim these building blocks as standalone innovations. Our contribution lies in the specific mechanisms connecting context, spectral conditioning, and dual-state aggregation.

First, cognitive and temporal context are incorporated into historical embeddings before spectral summarization. High-frequency magnitude statistics then generate input-dependent channel weights before GRU aggregation, separately for student and question histories. Context therefore shapes both the aggregated features and the spectral statistics governing their modulation.

Second, the prior-context mechanism combines concept-derived question representations with a learned modulation gate and a usage-threshold indicator. As defined in Eq. (8), prior injection is enabled only while the question’s cumulative training-usage count remains below �0��0�.

The controlled comparison above supports spectral conditioning, while Figure 5 demonstrates the prior module’s benefits across three other backbones. These results support the specific conditioning mechanism and the broader applicability of prior context.

W3 & Q2: Thanks for the suggestion. Our educational motivation draws on the distinction between durable learning and transient performance. Soderstrom and Bjork (2015) discuss their dissociation, while Corbett and Anderson’s knowledge-tracing model separates latent mastery from observed responses through slip and guess probabilities. These studies motivate distinguishing learning from performance, but do not establish a direct mapping between specific cognitive events and frequency bands.

Our spectral hypothesis concerns variation across interaction steps: brief deviations and abrupt transitions can contribute energy to higher frequencies, although their signatures generally span multiple bands. High-frequency statistics therefore provide cues to rapid variation rather than uniquely identifying slips or insights. Frequency here refers to the ordered interaction sequence; elapsed time is encoded separately.

FE performs frequency-conditioned channel reweighting, rather than directly amplifying selected high-frequency components. Higher-frequency content is not assumed to be inherently beneficial: its statistics condition a sigmoid gate over contextualized embeddings before GRU aggregation, with channel weights learned through predictive supervision. Our controlled comparison with non-spectral gates supports this design’s predictive usefulness, but does not by itself validate a correspondence between spectral components and specific cognitive processes.

References:

  1. Soderstrom, N. C., & Bjork, R. A. (2015). Learning versus performance: An integrative review. Perspectives on Psychological Science, 10(2), 176–199. https://doi.org/10.1177/1745691615569000
  2. Corbett, A. T., & Anderson, J. R. (1994). Knowledge tracing: Modeling the acquisition of procedural knowledge. User Modeling and User-Adapted Interaction, 4, 253–278. https://doi.org/10.1007/BF01099821

W4 & Q3: Thanks for the suggestion. As detailed in Appendix A.6, the LLM proposes prerequisite candidates using concept names and descriptions, without access to student responses or validation/test interactions. Three annotators review these candidates, remove unsupported relations, add missing relations, and resolve disagreements by voting. The resulting graphs encode dataset-specific prerequisite judgments; they are not assumed to be universally valid across curricula.

To assess sensitivity to imperfect prerequisites on ASSIST2017, we randomly removed 20%/50% of prerequisite edges or perturbed 10%/20%, retaining question–concept associations. We also evaluated a Q–C-only variant without prerequisite edges. Cold-Start Subset AUC is evaluated on the first 10 interactions of each question within the test split, while Overall AUC is evaluated on the full test split.

SettingCold-Start Subset AUC (%)Overall AUC (%)
Full Prior77.13±0.1281.66±0.05
Drop 20% Prerequisites76.72±0.1581.02±0.08
Drop 50% Prerequisites76.16±0.3980.29±0.31
Perturb 10% Prerequisites76.54±0.1781.04±0.09
Perturb 20% Prerequisites75.87±0.3381.15±0.23
Q–C Only75.69±0.1880.62±0.10
Strongest Baseline75.26±0.1680.35±0.14

These results support the usefulness of imperfect prerequisites on the evaluated question subset, while substantial edge removal limits overall gains and cross-curriculum transfer remains untested.

W5: Thanks for the suggestion. The gains are supported by consistent results, statistical significance, and cold-start performance.

DSCA-KT achieves the highest mean AUC across all ten dataset–setting combinations in Table 2, exceeding the strongest baseline in each setting by 1.14–2.86 percentage points, and achieves the highest mean AP in nine combinations. These AUC improvements are statistically significant under a two-sided paired t-test (p<0.05p < 0.05), as reported in Table 2.

On the ASSIST2017 subset comprising each question’s first 10 test interactions, DSCA-KT achieves 77.13% AUC versus 75.26% for the strongest baseline, a 1.87-point gain. Figure 5 additionally demonstrates benefits from transferring the prior module to three other backbones during early question interactions.

On ASSIST2012, training time increases from 67.852 to 70.637 s/epoch (+4.1%) relative to DyGKT, while inference time increases from 5.125 to 5.651 s (+10.3%). These measurements indicate that the predictive improvements are achieved with relatively limited additional runtime in the evaluated setting, alongside benefits when question histories are sparse.

W6 & Q4: Thanks for the suggestion. We distinguish three cold-start cases. New questions over existing concepts. Our prior-context mechanism directly targets this case by using associated concept representations when question histories are limited.

New students without interaction history. Empty histories are represented through zero-padding, while prediction can use available question-side representations and contextual information. As responses become available, the student’s historical representation updates chronologically, with model parameters fixed during evaluation. We additionally evaluated new-student predictions on ASSIST2017, grouping predictions by the number of student interactions observed strictly before the prediction timestamp; the zero-history group corresponds to students’ first interactions.

Student History LengthAUC (%)AP (%)
075.60±0.4867.10±0.35
1–277.15±0.3868.75±0.28
3–578.85±0.2970.55±0.20
6–1080.10±0.2072.10±0.12

DSCA-KT remains predictive without student history, with higher mean AUC and AP in groups with more observed interactions.

New concepts. A concept already represented in the knowledge graph can provide information even when the current student has never encountered it. A concept outside the model’s established vocabulary is a different case: when absent from both the training vocabulary and knowledge graph, the current mechanism does not guarantee an effective representation. This scenario remains an important direction for further investigation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.