Pith. sign in

REVIEW 4 major objections 4 minor 3 references

A Human-Centric Approach to Explainable AI for Personalized Education

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This thesis argues that post-hoc explainers are too inconsistent to serve as a trust layer in education, and that interpretable-by-design architectures plus an LLM explanation layer can deliver explanations that are accurate, real-time…

desk verdict A cumulative PhD thesis with a strong peer-reviewed core and a self-defined evaluation framework that needs scrutiny on criterion completeness. read the letter →

arxiv 2505.22541 v1 pith:YE6IUTVH submitted 2025-05-28 cs.LG cs.CY

Vinitra Swamy This is my paper
classification cs.LGcs.CY
keywords ExplainableAIHuman-CentricPersonalizededucationPost-hocexplainersInterpretable-by-designarchitecturesStudentsuccesspredictionCounterfactualexplanationsLLM-XAI
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that popular post-hoc explainers—LIME, KernelSHAP, PermSHAP, CEM, and DiCE—systematically disagree when applied to the same student success models, making them unsuitable as a basis for trust in educational AI. It argues that the field should shift to inherently interpretable model architectures, which can match non-interpretable baselines in predictive performance while offering faithful, real-time explanations. The thesis then adds an LLM-based communication layer that turns those explanations into natural-language feedback, which students preferred 89.2% of the time in a user study. A sympathetic reader would take the central claim to be that a human-centric XAI stack—intrinsic interpretability plus an LLM explanation layer—can make AI-driven personalized education both accurate and trustworthy enough for real classrooms.

What carries the argument

The load-bearing mechanism is a set of five design targets for explanations—accurate with certainty, real-time, consistent, actionable, and human-understandable—used as the evaluation yardstick throughout the thesis. The technical machinery consists of three complementary components: MultiModN, which sequentially fuses latent representations from multiple modalities and therefore makes each modality's contribution naturally inspectable; InterpretCC, which uses adaptive sparsity to activate only the features or feature groups needed for each prediction, so the explanation is the actual decision path; and iLLuMinaTE, an LLM-XAI pipeline whose multi-stage prompts align generated feedback with cognitive theories of explanation. Together these components are intended to satisfy the five design targets simultaneously, without sacrificing accuracy.

What would settle it

Give teachers and students the same predictions with explanations from both a post-hoc pipeline and the proposed intrinsic-plus-LLM pipeline, ask them to make real advising decisions under blind conditions, and include fairness, causal validity, and cost of acting among the evaluation criteria; if post-hoc explanations prove equally actionable or more trusted, the thesis's central ranking would not hold.

Watch

Extended reading notes

Core claim

The core discovery is that explainability method choice, not the underlying model or the student data, dominates what an explanation highlights: across five courses and five explainers, the feature importance distributions clustered by explainer rather than by course, with LIME consistently farthest from the others. This systematic disagreement, together with expert educators disagreeing about which explanations to trust, motivates the thesis's move to architectures that are interpretable by design. The proposed models—MultiModN, a sequential multimodal modular network, and InterpretCC, a sparse mixture-of-experts model—match non-interpretable baselines on balanced accuracy while providing explanations directly tied to the model's own decision process. Adversarial training improves the consistency of counterfactual explanations but not of LIME and SHAP, showing that consistency must be addressed holistically. Finally, the iLLuMinaTE pipeline uses large language models to convert feature-based explanations into concise, actionable feedback aligned with social-science explanation theories, and students prefer these explanations for clarity and actionability.

Load-bearing premise

The five design targets are the thesis's own yardstick, drawn from a viewpoint paper rather than from stakeholders, so the verdict that post-hoc methods fail and intrinsic models succeed presupposes that these criteria are the right ones.

Editorial extensions

If this is right

  • Classroom adoption of AI becomes more feasible because students receive instant, understandable feedback that they can act on, rather than raw feature scores.
  • Post-hoc explainers should not be treated as a trust layer in education, since their explanations vary more with method choice than with the actual model or data.
  • Interpretable-by-design architectures can match black-box accuracy while providing faithful, real-time explanations, removing the presumed tradeoff in this setting.
  • Adversarial training can stabilize counterfactual explanations, but LIME and SHAP remain sensitive to random perturbations, so explainer consistency cannot be fixed by model robustness alone.
  • Course design decisions can be grounded in explanations when educators are shown an ensemble of explainers, with a majority finding at least one actionable interpretation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the evaluation criteria were broadened to include fairness, causal validity, or the cost of acting on an explanation, the relative ranking of post-hoc versus intrinsic approaches could change; the thesis does not test those criteria.
  • The same evaluation template—five design targets plus user studies—could be exported to other human-centric domains such as healthcare or hiring, although the thesis only demonstrates it in education.
  • A testable extension is whether the LLM communication layer would make post-hoc explanations appear equally actionable, which would separate the communication win from the architecture win.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. This PhD-style manuscript argues that post-hoc explainability methods (LIME, KernelSHAP, PermSHAP, CEM, DiCE) are systematically inconsistent when applied to student success prediction models and are therefore unsuitable as a trust layer for personalized education. It then proposes a shift toward intrinsically interpretable architectures: MultiModN (sequential multimodal fusion), InterpretCC (sparse mixture-of-experts with feature or group routing), adversarial training for explainer stability, and iLLuMinaTE (an LLM-based theory-driven communication layer that turns XAI outputs into feedback for students). The thesis evaluates these contributions across large MOOC datasets, five explainers, and four human studies with professors, teachers, learning scientists, and students, claiming that explainability can be made accurate, real-time, consistent, actionable, and human-understandable.

Significance. The constituent chapters of this thesis have already passed peer review at strong venues (NeurIPS, ICLR, AAAI, LAK, Learning@Scale, EDM), and the visible methodology is careful: balanced accuracy on imbalanced MOOCs, stratified sampling of representative students, five random seeds with overlapping confidence intervals in Chapter 8, and multiple human studies with domain experts. The authors also release code and models for the main technical contributions. If the framework-level claims are taken at face value, the work would be an important step toward human-centric XAI in education, particularly the move from post-hoc approximation to interpretable-by-design architectures combined with an LLM communication layer. The main risk is not the individual experiments but the framework that ties them together: the five design targets are sourced from the author's own under-review viewpoint paper and are then used as the yardstick for all comparative verdicts. This circularity is load-bearing and needs to be addressed head-on.

major comments (4)
  1. [Section 1.3 and Chapter 10] The five design targets (accurate with certainty, real-time, consistent, actionable, human-understandable) are introduced in Section 1.3 as 'motivated originally in [SFK23]', the author's own under-review viewpoint paper, rather than as a standard elicited from stakeholders or derived from an established framework. Every contribution chapter then demonstrates that its method satisfies these targets, and Chapter 10 organizes the conclusions around them. This creates a circularity at the framework level: the comparative verdicts (post-hoc explainers fail; intrinsic models plus iLLuMinaTE succeed) presuppose that this self-authored yardstick is correct, complete, and correctly weighted. The user studies in Chapter 7 (Table 7.4) and Chapter 9 (Figure 9.5) let participants rank or choose within the dimensions the authors made available, but they do not elicit the dimensions themselves or their relative weights. If real deployment requires additional or differently weighted criteria, such as fairness, causal validity, the cost of acting on an explanation, privacy, or downstream learning outcomes, the evidence as presented cannot rank the proposed methods against those criteria. The thesis is transparent about the source of the framework, but transparency does not supply the missing stakeholder validation. I would ask for either a stakeholder elicitation study that derives and weights the design targets independently, or a substantial reframing that treats the five targets as working hypotheses rather than as the established basis for comparative conclusions.
  2. [Section 3.2.2 and Figure 3.4] The early-dropout filtering step removes students whose predicted failure probability exceeds 0.99 according to a logistic regression trained on assignment grades from the first two weeks. This is an ad hoc preprocessing choice: the threshold is tuned by grid search on the training courses, and the transfer results reported in Tables 3.2 and 3.3 and in Figures 3.3 are computed on the filtered population. Figure 3.4 shows that the broad patterns are similar on the full population, and the text states that the predictors 'have high accuracy on these students as well', but no quantitative comparison of transfer performance on filtered versus unfiltered populations is provided, nor is there a sensitivity analysis of the threshold. Because the filtering disproportionately affects courses with different assignment schedules, lengths, and dropout rates, this could bias the conclusion that behavior-plus-meta models transfer better than same-course baselines. The claim of generalizable student modeling needs a robustness check showing that the transfer rankings are stable across reasonable choices of the filtering threshold and across courses with very different early-assessment structures.
  3. [Section 4.3.2 and Figure 4.3] The central 'systematic disagreement' result is based on explanations for 100 uniformly sampled students per course and a single trained BiLSTM per course in Chapter 4, with a related design in Chapter 5. The sampling strategy balances on predicted probability intervals, but the resulting set is not shown to be representative of the full student distribution, and the PCA, Spearman correlation, and Jensen-Shannon distance analyses do not report bootstrap confidence intervals or variation across random seeds. Chapter 8 later studies seed variation and finds substantial explainer variance, but the earlier chapters' strong claim that the choice of explainer dominates the choice of course or model would be considerably more robust if the original comparison included multiple model initializations and repeated sampling with uncertainty estimates. As it stands, the claim could be sensitive to the particular 100-student samples and the single trained model per course.
  4. [Chapter 9 and overall framing] The 'actionable' design target is validated through student self-reported preference and Likert ratings (e.g., 89.2% preference, usefulness and actionability scores in Figures 9.5 and 9.6), not by measuring whether students or teachers can actually implement the suggestions or whether doing so improves learning outcomes. The abstract's claim that the work 'lays a foundation' for human-centric AI systems is more modest than the introduction's suggestion that classroom adoption becomes feasible, but the gap remains: perceived actionability is demonstrated, realized actionability is not. A small deployment study, a follow-up behavior measurement, or an explicit limitation statement that realized actionability is outside the scope would help calibrate the conclusions.
minor comments (4)
  1. [Section 2.2] The text says 'As illustrated in Figure 8.1' when describing the taxonomy of explainability stages; the correct cross-reference should be Figure 2.1.
  2. [Sections 1.2 and 2.2] There are typos: 'innaccuracies' in Section 1.2 and 'vase majority' in Section 2.2. Also, the sentence 'most of these studies have performed a-posteriori or post-hoc analyses, and not models for early success prediction' in Section 2.1 is awkward and should be rephrased.
  3. [Tables 4.1 and 5.3] Both tables refer to 'Section 7.5' for the features identified as important by at least one explainability method, but the relevant analyses appear in Sections 4.3.3 and 5.3; these cross-references should be corrected.
  4. [Table 9.1] Table 9.1 reports that explanations are 'Annotated by experts and GPT-4o', but the number of expert annotators and their inter-rater agreement are not stated in the main text; the reader should be pointed to Appendix E.2 with a concrete agreement figure.

Circularity Check

0 steps flagged · score 2.0 of 10

No material circularity: the only self-referential element is the [SFK23]-motivated five-target framework in Section 1.3, which frames but does not by construction force the thesis's empirical claims.

full rationale

The closest thing to a circular step is Section 1.3, where the thesis states 'we define five design targets that explanations should fulfill, motivated originally in [SFK23]' — a viewpoint paper co-authored by the thesis author. This is a minor self-citation and it does supply the normative yardstick (accurate-with-certainty, real-time, consistent, actionable, human-understandable) used to organize the contributions. However, the paper's load-bearing results are not equivalent to that yardstick by construction. The claim that post-hoc explainers are inconsistent is supported by direct quantitative comparisons (Spearman correlation, Jensen-Shannon distance, PCA) across five explainers on independently trained BiLSTM models (Chapter 4), by educator interviews (Chapter 5), and by random-seed and adversarial-robustness experiments (Chapter 8). The claims that the proposed interpretable architectures match non-interpretable baselines rest on performance comparisons with reported confidence intervals (Chapters 6 and 7), and the iLLuMinaTE preference and actionability claim rests on a 114-student user study measuring choices and Likert ratings (Chapter 9). The five targets are a normative scheme, not an equation into which findings are substituted; choosing the scheme affects how findings are weighted but does not make the measurements equal to the scheme's assumptions. The GPT-4o annotation of LLM-generated explanations (Table 9.1) is a measurement-protocol concern rather than a derivation-level circularity, and the thesis also reports expert agreement (Table E.2). I therefore find no significant circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claims rest on a small set of modeling choices and domain assumptions. The early-dropout filter and the 100-student explainer sample shape all Part I results; the five design targets, sourced from the author's own viewpoint paper, shape every Part II evaluation; and the MVK21 behavioral features are inherited from prior work without revalidation. The iLLuMinaTE pipeline depends on hand-written theory prompts. No new physical or conceptual entities are postulated, so invented_entities is empty. Data and code release are partial: code for Chapters 3 and 4 is linked, while the MOOC corpus and user pools are not.

free parameters (4)
  • Early-dropout filtering threshold = P(fail) > 0.99
    Students with predicted failure probability above the threshold are removed before training and evaluation in Chapters 3 and 4; the threshold is selected by grid search over {0.96,...,0.999} to maximize balanced accuracy, and all transfer, explainer, and consistency results are computed on this filtered population.
  • Explainer evaluation sample = 100 students per course (50 pass, 50 fail)
    Chapter 4 selects 100 representative students per course by uniform sampling in predicted-probability order, stratified by class, to keep explainer computation feasible; every cross-explainer comparison in Chapters 4 and 5 is computed on this sample.
  • InterpretCC sparsity controls = Gumbel tau and activation threshold tuned per dataset
    Appendix D.4 reports sensitivity experiments for the Gumbel-softmax temperature and activation threshold; the values are chosen per dataset by the authors and control how many features or groups each explanation contains.
  • iLLuMinaTE prompt templates = Six theory prompt designs (Appendix E.2)
    The LLM explanation layer is driven by researcher-written prompts implementing six social-science theories; the actionability and preference results are tied to these specific hand-written templates.
assumptions (5)
  • domain assumption The five design targets (accurate with certainty, real-time, consistent, actionable, human-understandable) are the correct normative criteria for evaluating XAI in education.
    Stated in Section 1.3 as 'motivated originally in [SFK23]', the author's own under-review viewpoint paper; all method chapters are evaluated against these self-defined targets.
  • domain assumption The four behavioral feature sets from [MVK21] (Regularity, Engagement, Control, Participation, 42-45 features) sufficiently capture student behavior for success prediction and explanation.
    Chapters 3, 4, 5, 8, and 9 build all student models on these inherited features without revalidation in the thesis.
  • ad hoc to paper Removing early-dropout students (predicted failure probability above 0.99) does not bias the comparisons of transfer models and explainers.
    The filter is introduced by the authors in Sections 3.2.2 and 4.2.2; no robustness analysis shows that the explainer-disagreement and transfer conclusions are invariant to the chosen threshold.
  • domain assumption Social-science theories of explanation (contrastive, Pearl's model, necessity and robustness, and others) can be faithfully operationalized by prompting an LLM.
    Chapter 9's iLLuMinaTE assumes the hand-written prompts implement the theories correctly; alignment is verified by human and GPT-4o annotation with partial agreement.
  • domain assumption Self-reported preference and perceived actionability in lab or online user studies predict real classroom use and learning benefit.
    The human-centric claims of Chapters 5, 7, 8, and 9 rest on Likert ratings, rankings, and interview coding rather than measured behavior change in deployed systems.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Human-Centric Approach to Explainable AI for Personalized Education." pith.science (2026). https://pith.science/paper/YE6IUTVH

@misc{pith2026250522541,
  author       = {Pith},
  title        = {Pith review of: A Human-Centric Approach to Explainable AI for Personalized Education},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YE6IUTVH}},
  note         = {Machine review of arXiv:2505.22541}
}
read the original abstract

Deep neural networks form the backbone of artificial intelligence research, with potential to transform the human experience in areas ranging from autonomous driving to personal assistants, healthcare to education. However, their integration into the daily routines of real-world classrooms remains limited. It is not yet common for a teacher to assign students individualized homework targeting their specific weaknesses, provide students with instant feedback, or simulate student responses to a new exam question. While these models excel in predictive performance, this lack of adoption can be attributed to a significant weakness: the lack of explainability of model decisions, leading to a lack of trust from students, parents, and teachers. This thesis aims to bring human needs to the forefront of eXplainable AI (XAI) research, grounded in the concrete use case of personalized learning and teaching. We frame the contributions along two verticals: technical advances in XAI and their aligned human studies. We investigate explainability in AI for education, revealing systematic disagreements between post-hoc explainers and identifying a need for inherently interpretable model architectures. We propose four novel technical contributions in interpretability with a multimodal modular architecture (MultiModN), an interpretable mixture-of-experts model (InterpretCC), adversarial training for explainer stability, and a theory-driven LLM-XAI framework to present explanations to students (iLLuMinaTE), which we evaluate in diverse settings with professors, teachers, learning scientists, and university students. By combining empirical evaluations of existing explainers with novel architectural designs and human studies, our work lays a foundation for human-centric AI systems that balance state-of-the-art performance with built-in transparency and trust.

Figures

Figures reproduced from arXiv: 2505.22541 by the authors.

Figure 2.1
Figure 2.1. Explainability can be post-hoc (e.g., LIME, SHAP), in-hoc (e.g., saliency methods), [PITH_FULL_IMAGE:figures/full_fig_p037_2_1.png] view at source ↗
Figure 3.1
Figure 3.1. Our transfer approach based on behavior and course metadata features, from logs [PITH_FULL_IMAGE:figures/full_fig_p044_3_1.png] view at source ↗
Figure 3.2
Figure 3.2. Model architectures adopted in our study. [PITH_FULL_IMAGE:figures/full_fig_p048_3_2.png] view at source ↗
Figures from the paper (41 more)
Figure 3.3
Figure 3.3. Figure 3.3: Comparison of the balanced accuracy achieved by our behavior only and behavior [PITH_FULL_IMAGE:figures/full_fig_p053_3_3.png]
Figure 3.4
Figure 3.4. Figure 3.4: Average performance of BO, BSM, BTM across transfer courses with and without previous iterations, both on the filtered (w/ early-dropout students) and the entire student population. that pass in meta features at different locations of the model (see Section 3.2.2). W…
Figure 3.5
Figure 3.5. Figure 3.5: Ablation study on the meta features for the [PITH_FULL_IMAGE:figures/full_fig_p056_3_5.png]
Figure 3.6
Figure 3.6. Figure 3.6: Importance of the different meta features on the predictive power of the [PITH_FULL_IMAGE:figures/full_fig_p056_3_6.png]
Figure 3.7
Figure 3.7. Figure 3.7: Comparative balanced accuracy of different fine-tuned meta models. [PITH_FULL_IMAGE:figures/full_fig_p058_3_7.png]
Figure 4.1
Figure 4.1. Figure 4.1: Our experimental pipeline, from data processing to post-hoc explainability meth [PITH_FULL_IMAGE:figures/full_fig_p062_4_1.png]
Figure 4.2
Figure 4.2. Figure 4.2: Heatmap of normalized feature importance scores (log scale) across explainability [PITH_FULL_IMAGE:figures/full_fig_p069_4_2.png]
Figure 4.3
Figure 4.3. Figure 4.3: PCA of feature importance scores for five explainability methods across five courses. [PITH_FULL_IMAGE:figures/full_fig_p071_4_3.png]
Figure 4.4
Figure 4.4. Figure 4.4: Comparison of feature importance scores across courses using Spearman’s Rank [PITH_FULL_IMAGE:figures/full_fig_p072_4_4.png]
Figure 4.5
Figure 4.5. Figure 4.5: Comparison of feature importance scores across courses using Jensen-Shannon [PITH_FULL_IMAGE:figures/full_fig_p072_4_5.png]
Figure 4.6
Figure 4.6. Figure 4.6: Prerequisite skill structure for DSP 1. ing Jensen-Shannon Distance. Larger numbers represent higher dissimilarity. The Jensen￾Shannon Distance heatmaps confirm the observations made using Spearman’s Rank-Order Correlation (see [PITH_FULL_IMAGE:figures/full_fig_p073…
Figure 4.7
Figure 4.7. Figure 4.7: Importance scores for LIME, PermSHAP, and CEM for week 5. [PITH_FULL_IMAGE:figures/full_fig_p075_4_7.png]
Figure 4.8
Figure 4.8. Figure 4.8: Importance scores for LIME, PermSHAP, and CEM for week 9. [PITH_FULL_IMAGE:figures/full_fig_p075_4_8.png]
Figure 5.1
Figure 5.1. Figure 5.1: Our framework involves four steps: 1) extracting clickstream data from five pairs [PITH_FULL_IMAGE:figures/full_fig_p081_5_1.png]
Figure 5.2
Figure 5.2. Figure 5.2: Jaccard Similarity and Spearman’s Rank-Order Correlation between XAI methods ordered by method. then uniformly sampled 50 failing students from this ordered interval. We repeated this exact same procedure to sample the 50 passing students. This sampling procedure ens…
Figure 5.3
Figure 5.3. Figure 5.3: Flipped vs. MOOC Functional Programming course; showcasing absolute, normal [PITH_FULL_IMAGE:figures/full_fig_p091_5_3.png]
Figure 5.4
Figure 5.4. Figure 5.4: Low vs. High active learning: Geo and AN2 course; showcasing absolute, normal￾ized importance scores across weeks. MOOC setting, LIME chooses primarily problem features as important (notably num-submit￾quiz for failure and distinct-probs-quiz for success), but in the…
Figure 6.1
Figure 6.1. Figure 6.1: Comparison of modular MultiModN (a) vs. monolithic P-Fusion (b). MultiModN inputs any number/combination of modalities (mod) into a sequence of mod-specific en￾coders (e). It can skip over missing modalities. A state (s) is passed to the subsequent encoder and update…
Figure 6.2
Figure 6.2. Figure 6.2: MultiModN does not compromise performance in single-tasks. AUROC for six binary prediction tasks in (a) MIMIC, (b) EDU, and (c) Weather2k. Tasks predicted by P-Fusion are compared with MultiModN. 95% CIs are shaded. Model evaluation and metrics. All results represent…
Figure 6.3
Figure 6.3. Figure 6.3: Multi-task MultiModN maintains baseline performance in individual tasks. Single- and multi-task MultiModN on the prediction of individual tasks, compared with the monolithic P-Fusion (can only be single-task). AUC for binary (left) and MSE for continuous (right). Err…
Figure 6.4
Figure 6.4. Figure 6.4: Inherent modality-specific model explainability in MultiModN. Heatmaps show individual modality contributions (IMC) (top) and cumulative contributions (CP) (bottom): respectively importance score (global explainability) or cumulative probability (local ex￾plainabilit…
Figure 6.5
Figure 6.5. Figure 6.5: MultiModN is robust to catastrophic MNAR failure. Impact of MNAR missingess on MultiModN vs. P-Fusion. Both models are trained on four versions of the MIMIC dataset with 0—80% MNAR. They are then tested on either a test set with no MNAR missingness (- - -) or a test …
Figure 7.1
Figure 7.1. Figure 7.1: InterpretCC Architectures: Feature Gating (left, individual features): (i) All fea￾tures are input into a discriminator network that outputs a sparse feature activation mask; (ii) Only the features selected via the mask are passed to a predictive network for the fina…
Figure 7.2
Figure 7.2. Figure 7.2: InterpretCC Feature Gat￾ing Sparsity: % of features activated per data point across five representa￾tive datasets. change required to do so would no longer respect the original model designIII; instead we report text results for FRESH. Additional sensitivity and arch…
Figure 7.3
Figure 7.3. Figure 7.3: AG News and SST: # of ICC subnetwork activations (left) vs. avg. activation weights (right), grouped by subnetworks based on the Dewey Decimal Code. Model Relationship to Underlying Data Patterns Relationship to Model Ground Truth Alignment (GTA) Rank Agreement (RA) …
Figure 7.5
Figure 7.5. Figure 7.5: Tukey’s Honest Significant Differ￾ence (HSD) Test for each user study criterion and overall global satisfaction. ICC FG (with highest overall satisfaction in [PITH_FULL_IMAGE:figures/full_fig_p127_7_5.png]
Figure 8.1
Figure 8.1. Figure 8.1: Study Overview: We begin by consulting learning scientists to assess their percep￾tions of explanation variance. In the baseline pathway (orange), we 1) extract clickstream data and 2) train a model for student success. In the robustness pathway (blue), we 1) generat…
Figure 8.2
Figure 8.2. Figure 8.2: Learning Science Expert Study: Three learning scenarios at different granularities of instruction (principal, tutor, teacher) presented to learning science experts. The interviewer asks the expert to specify which needs of explainable AI are most relevant in each sce…
Figure 8.3
Figure 8.3. Figure 8.3: Learning Science Expert Study: Screenshots of student scenario 1 for the TA context (left) and student scenario 3 for the models context (right). or two computational models (Model context). Below we describe each context in more detail. First, in the TA context, the…
Figure 8.4
Figure 8.4. Figure 8.4: Coded analysis of expert interviews: Average number of IS (TAs or Models) found trustworthy (left), number of features selected for intervention (middle), and percentage of interviewees who trusted at least one IS (right). Next, we investigated the differences betwee…
Figure 8.5
Figure 8.5. Figure 8.5: Distribution of the standard deviations of feature importance scores across five [PITH_FULL_IMAGE:figures/full_fig_p144_8_5.png]
Figure 8.6
Figure 8.6. Figure 8.6: Feature important scores of the top features for five randomly initialized models on [PITH_FULL_IMAGE:figures/full_fig_p144_8_6.png]
Figure 8.7
Figure 8.7. Figure 8.7: Percentage of features found important by CEM over three approaches (baseline [PITH_FULL_IMAGE:figures/full_fig_p146_8_7.png]
Figure 8.8
Figure 8.8. Figure 8.8: Temporal analysis of the cumulative average importance scores for four courses [PITH_FULL_IMAGE:figures/full_fig_p146_8_8.png]
Figure 8.9
Figure 8.9. Figure 8.9: Distribution of the standard deviations of feature importance scores across five [PITH_FULL_IMAGE:figures/full_fig_p147_8_9.png]
Figure 8.10
Figure 8.10. Figure 8.10: Sources of Explainer Inconsistency, measured by distributions of the std. devia￾tions of importance scores, across CEM, LIME, and SHAP explainers for 2 flipped classroom courses (Linear Algebra and FP) and 4 MOOC courses. Mitigation Strategies. We extend to our anal…
Figure 8.11
Figure 8.11. Figure 8.11: Explanation Variability across FGSM, PGD and Baseline methods measured by distributions of the std. deviations of importance scores, across CEM, LIME, and SHAP explainers for 2 flipped classroom courses (Linear Algebra and FP) and 4 MOOC courses. approximation like …
Figure 9.1
Figure 9.1. Figure 9.1: iLLuMinaTE involves four steps: 1) modeling course data for student success prediction, 2) using XAI methods to extract feature importance scores, 3) selecting important aspects of the explanation through an LLM aligned with a given social science theory, 4) presenti…
Figure 9.4
Figure 9.4. Figure 9.4: Readability metrics. Flesch Kincaid, Gunning Fog, SMOG Index, Grammar Issues across LLM (blue), course (green), and explainer (orange). Lower scores are better. Model Explainer Explanation Selection Explanation Presentation Gemma2 9b CEM 0.941 ± 0.202 0.791 ± 0.212 L…
Figure 9.3
Figure 9.3. Figure 9.3: Explanation quality by LLM. Degree of instruction-following for expla￾nations generated by GPT-4o, Gemma2 9b, and Llama3 70b. Average (with standard deviation) of “Yes” answers, annotated by GPT-4o. In a second analysis, we compared different LLMs’ abilities to gener…
Figure 9.5
Figure 9.5. Figure 9.5: Student preference of presented explanations for passing (left) and failing (right) student predictions. Percentage of times a student chose each method when it was available. Higher scores are better. Usefulness Actionability Base RS AC NR CoT Pearl Con 1 2 3 4 5 Ba…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

3 extracted references · 1 linked inside Pith

  1. [146]

    How to Quantify Student’ s Regularity?

    ISSN : 2307-387X. [Bor+16a] Mina Shirvani Boroujeni et al. “How to Quantify Student’ s Regularity?” In:Adap- tive and Adaptable Learning. Ed. by Katrien Verbert, Mike Sharples, and Tomaž Klobuˇ car. 2016. [Bor+16b] Mina Shirvani Boroujeni et al. “How to quantify student’ s regularity?” In:ECTEL. Springer. 2016, pp. 277–291. [BP15] Journal Broadbent and W ...

    arXiv 2015

  2. [405]

    Generalized additive models

    DOI : 10.1109/CCWC54503.2022.9720806. [Has17] Trevor J Hastie. “Generalized additive models”. In:Statistical models in S. Rout- ledge, 2017. [Haz+21] Hussein Hazimeh et al. “Dselect-k: Differentiable selection in the mixture of experts with applications to multi-task learning”. In: Advances in Neural Infor- mation Processing Systems 34 (2021), pp. 29335–2...

    arXiv 2021

  3. [2023]

    A survey of transformers

    URL : https://openreview.net/forum?id=i2_TvOFmEml. [Lin+22] Tianyang Lin et al. “A survey of transformers”. In: AI open 3 (2022), pp. 111–132. 257 Chapter E BIBLIOGRAPHY [Lin91] Jianhua Lin. “Divergence measures based on the Shannon entropy”. In: IEEE Tran. on Info. Theory 37.1 (1991), pp. 145–151. [Lip90] Peter Lipton. “Contrastive explanation”. In:Royal...

    arXiv 2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.