Shared Affect Prototypes for Federated Knowledge Tracing

Tutoring software that tracks only whether a student answers correctly misses why the student is struggling. Frustration, boredom and confusion change what help is useful next. A sensorless line of work shows that much of this can be inferred from ordinary interaction logs: response times, attempt counts and gaps between questions. DASKT clusters such log features into four affective states and feeds them into a knowledge-tracing model, which predicts a student's future correctness from their past answers. It reports a large gain on ASSISTchall (AUC 0.8529 against 0.8074 for the strongest prior model, DIMKT). It does this centrally, by pooling every school's timing logs. Those timing logs are exactly what schools are reluctant to export. Federated learning, where each school trains locally and shares only model updates, has reached educational modeling in FedCD, but FedCD models correctness alone and ignores affect.

The obvious move is to federate DASKT. This proposal argues that the naive version is broken in a specific way. DASKT's affect states are cluster labels defined relative to population statistics of latency and effort. A "slow" response in one school or for one student is "fast" for another. When each school clusters locally, "frustration" in school A and "frustration" in school B are different regions of feature space, and they can even swap labels entirely. Averaging the downstream networks then mixes incompatible meanings. Our claim is that affect can be made consistent across schools without pooling logs by doing two things together. First, each learner's behavioral features are normalized against that learner's own running baseline, on their own device or school server. Second, a small set of affect prototypes, meaning cluster centers in this normalized space, is learned with federated clustering and shared by everyone. This mirrors the strongest empirical finding in the affective-tutoring literature: emotional expression is idiosyncratic, and generic detectors collapse on unseen learners (Alyüz et al. report 63.41% F1 for generic versus 90.89% for personalized log-based models in problem solving).

Core Claim

Affect-aware knowledge tracing can keep most of its centralized accuracy gain under federation, and keep it fairly across schools, only if affect is defined relative to each learner's own behavior and anchored to globally shared prototypes. Federating the model while leaving affect definitions school-local, or defining them against a global population scale, should lose much of the gain. If the per-learner baseline is the operative ingredient, an ablation that removes it should recover the naive federated failure.

Research Program

The model follows DASKT's pipeline but changes what is local and what is shared. For learner $u$ at step $t$, the raw behavioral vector $x_{u,t}$ contains three kinds of features: confidence (correctness relative to response time), interest (engagement frequency) and effort (attempts, inter-question interval). It is normalized against the learner's own exponentially weighted history:

$$z_{u,t} = (x_{u,t} - \mu_{u,t}) / \sigma_{u,t},$$

where $\mu_{u,t}$ and $\sigma_{u,t}$ are that learner's running mean and standard deviation, and these never leave the client. Each step is softly assigned to $K=4$ shared prototypes $c_1,\dots,c_K$ by distance in $z$-space. The prototypes are updated by federated k-means: clients send per-prototype sums and counts through secure aggregation, and the server divides. For a cold-start learner, the running baseline is initialized from the school-level statistics until enough history accumulates. The soft assignment feeds DASKT's temporal attention component and its affect-fused LSTM.

Parameter decoupling follows FedCD's logic. Exercise embeddings, the affect-transition attention module and the LSTM are shared and aggregated. Student state, the normalization statistics and the raw logs stay local. We adopt FedCD's loss-weighted aggregation, which gives schools with higher prediction loss more weight, because FedCD shows that volume-weighted averaging favors high-performing schools.

The analytic contribution is a direct measurement of prototype drift. We will quantify how far locally clustered affect centers diverge across schools under each normalization scheme, and how that divergence predicts the per-school loss of AUC. This turns "non-IID data hurts" into a concrete, testable statement about affect semantics.

Baselines

The naive reference point is local-only training: each school trains plain knowledge tracing (DKT) and the DASKT variant on its own data, with no sharing. This shows whether collaboration helps at all and how small schools suffer.

The centralized upper bound is DASKT trained on pooled data, as in DASKT. It reports AUC 0.8529 on ASSISTchall. We will reproduce this number on our splits before any comparison, since our school-based partitioning differs from theirs. Success means closing most of the gap between local-only and this bound.

The strongest existing federated competitor is a FedCD-style decoupled model without affect: shared exercise embeddings, local student states and loss-weighted aggregation, applied to a knowledge-tracing backbone. FedCD reports 72.01% diagnostic accuracy against 68.16% for FedAvg on ASSIST2009, with group fairness improving from 0.167 to 0.084 (FedCD). That is a cognitive-diagnosis task, not sequential knowledge tracing, so the figures fix the expected size of the fairness effect rather than serving as a direct target. We found no reported result on Lacuna for federated affect-aware knowledge tracing.

The naive federated affect model is DASKT under FedAvg with school-local k-means. The key ablations are federated prototypes over globally standardized features, and federated prototypes over per-school standardized features. Both isolate whether the per-learner baseline matters, as opposed to merely sharing prototypes.

For fairness, every method gets the same backbone size, the same school partition, the same number of communication rounds and local epochs, and an equal hyperparameter search budget, tuned on held-out students within training schools. The proposal beats the strongest competitor if it improves mean AUC over the FedCD-style model, and also improves the worst-school AUC, with significance assessed by paired bootstrap over students.

Why not just…?

Why not just run DASKT under FedAvg? Its affect states are school-local cluster labels with no shared meaning, so averaging the networks that consume them mixes incompatible signals. This proposal tests that claim rather than assuming it, and the naive federated DASKT is a baseline. If naive federation already matches centralized DASKT, the motivation fails and we will report that outcome.

Why not drop affect and federate plain knowledge tracing with FedCD-style decoupling? This is the real threat. Affect's centralized gain may partly reflect richer timing features rather than affect per se, and under federation a well-decoupled cognitive model may capture the same signal. The ablation that feeds raw normalized timing features directly, with no prototypes, will settle whether the affect bottleneck adds anything beyond the features. If it does not, the honest conclusion is that federated knowledge tracing needs per-learner timing normalization, not affect states. We will also check whether the prototypes still matter for interpretability, by measuring agreement with ASSIST2012's affect labels.

Why not just fine-tune a global model locally per student? Personalization is part of the answer, as Alyüz et al. show. However, individual students have too few frustrated or bored steps to fit affect locally. Per-learner normalization is a cheap, parameter-free form of personalization that leaves the scarce affect structure to the shared prototypes. We include a global-model-plus-local-fine-tuning baseline under the same budget to check this.

Why not add cameras or microphones and federated modality recovery? FedDISC reconstructs missing modalities for conversational emotion recognition with diffusion models. However, it was not evaluated in tutoring, its iterative denoising is costly at inference, and no large tutoring corpus with video is partitioned by school. DASKT admits that logs miss internal affect that never shows in behavior. Instruction-only periods, which Alyüz et al. find are best read from appearance, remain outside this proposal's scope.

Proposed Experiments

We will use ASSIST2009, ASSIST2012 and ASSISTchall, partitioning students into clients by school identifier, or by teacher or class where school identifiers are missing. This produces naturally non-identical clients of very different sizes. We hold out whole students for testing. A second split holds out entire schools to measure cold-start transfer.

We report AUC, both overall and for the worst school, and FedCD's group-fairness gap. On ASSIST2012 we also measure agreement between prototype assignments and the dataset's affect labels, comparing against DASKT's reported boredom consistency of over 80%. Prototype drift is measured as mean matched-centroid distance across schools.

The decisive check has three parts. Per-learner normalization with shared prototypes should beat both the school-local and the globally standardized variants on worst-school AUC. Its affect agreement on ASSIST2012 should stay near the centralized level. Its advantage over the no-affect FedCD-style model should exceed the bootstrap noise. If the global-standardization ablation does as well, the per-learner baseline is not the key ingredient, and the claim narrows to "shared prototypes suffice."

As a secondary privacy probe, in the spirit of the attribute-inference attacks on federated speech emotion models (Attribute Inference Attack of Speech Emotion Recognition in Federated Learning Settings), we train an attacker to infer school identity from the shared prototype statistics. The comparison is against the same attacker run on shared raw-feature aggregates.