Source-linked AI summary
An Incremental Knowledge-Graph Learner Digital Twin for Evidence-Grounded Early Warning
Ngoc Luyen Le, Marie-Hélène Abel, Bertrand Laforge
TL;DR
Early-warning alerts can indicate learner risk without identifying the learning evidence behind it. This paper routes prediction through an incremental knowledge-graph Learner Digital Twin that links features, attributions, alerts, evidence, and raw records. On OULAD, graph-mediated features preserve the predictive behavior of equivalent raw-table features across model families and checkpoints, while alerts can be traced through evidence chains.
Problem
Early-warning alerts may lack inspectable links from a risk signal to the learning traces, assessment events, or temporal patterns that justify it.
Source: §1 · p. 2Method
An incremental Learner Digital Twin replays traces and assessment records at temporal checkpoints, extracts features through graph queries, and represents predictions and evidence as linked graph entities.
Source: §1 · pp. 2–3, §3.1 · p. 5Results
Graph-mediated feature extraction preserves the predictive behavior of equivalent raw-table features across model families and checkpoints, and generated alerts can be traced to attributions, evidence items, and raw records.
Source: §5 · p. 13, §6 · p. 14Takeaways & Limitations
The study demonstrates that early-warning prediction can use a semantic, provenance-aware representation while preserving predictive performance and making alerts evidence-grounded and traceable.
Source: §1 · p. 2, §6 · p. 14The evaluation replays a historical dataset at predefined checkpoints and does not assess live ingestion, real-time integration, or instructor feedback loops.
Source: §5 · p. 13
Abstract
from arXiv · showhide
Early-warning models in learning analytics can identify learners at risk of failure or withdrawal, but the resulting alerts are often only weakly connected to inspectable learning evidence. A risk score may indicate that intervention is needed without showing which traces, assessment events, or temporal patterns justify the alert. This paper introduces an incremental Learner Digital Twin for evidence-grounded early warning, implemented as a lightweight knowledge graph. The twin is updated at temporal checkpoints by replaying learning traces and assessment records, and it serves both as a feature-extraction substrate and as an explanation-grounding substrate. Predictive features are derived through graph-mediated queries, while features, attributions, alerts, evidence items, and raw records are represented as queryable graph entities. The framework is instantiated and evaluated on a public learning-analytics benchmark under a temporal protocol. The results show that graph-mediated feature extraction preserves the predictive behavior of equivalent raw-table features across model families and checkpoints. Generated alerts are materialized as evidence chains linking risk decisions to feature attributions, evidence items, and raw platform records. The study shows that early-warning prediction can be routed through a semantic and provenance-aware representation while preserving predictive performance and adding traceability, verifiability, and evidence-grounded explanation.
1 Introduction
Early-warning systems can predict learner risk, but instructors may lack the concrete learning evidence behind an alert. The paper introduces an incremental knowledge-graph Learner Digital Twin that supports prediction and evidence-grounded, traceable alerts.
1 Introduction
- Feature-level explanations can name an engineered variable without identifying the learning events or raw records that substantiate an alert.The paper frames this as a gap between predictive performance and evidence-grounded interpretability.Source: §1 · p. 2
- The incremental Learner Digital Twin replays timestamped traces and assessment data at temporal checkpoints, then supplies graph-mediated features to an early-warning classifier.Predictions, attributions, alerts, and evidence items are materialized in the graph and linked to raw records; SHACL checks graph integrity.Source: §1 · pp. 2–3
- The paper's objective is to preserve predictive performance while routing early-warning prediction through a semantic, provenance-aware representation.Its stated aim is not to improve predictive accuracy.Source: §1 · pp. 2–3
2 Related Work
Prior work addresses early-warning prediction, learner representations, semantic educational graphs, and data provenance, but these strands are not commonly integrated into traceable, evidence-grounded alerts.
2 Related Work
- Early-warning research has largely emphasized predictive performance, timeliness, and deployment feasibility, while feature attributions may leave educational evidence uninspectable.The related-work discussion notes that instructors may need to inspect accessed resources, missed assessments, and supporting raw records.Source: §2 · p. 3
- Learner models often represent learner states or profiles, but are less often time-indexed, recurrently synchronized with platform traces, and connected to predictive alert generation.The paper uses a restricted digital-twin sense: a checkpoint-indexed learner representation synchronized by replayed observations and used for prediction and evidence-grounded explanation.Source: §2 · pp. 3–4
- Educational knowledge graphs commonly represent learners, resources, activities, or competencies rather than the internal objects of predictive early-warning pipelines.Provenance and SHACL offer mechanisms for linking outputs to sources and validating graph constraints, yet individual alerts remain weakly represented as jointly queryable evidence chains.Source: §2 · p. 4
3 Methodology
The methodology defines a checkpoint-indexed Learner Digital Twin that extracts predictions from temporally available learning data and preserves queryable links from alerts to their supporting records.
3.1 Problem Formulation
- At each checkpoint, the system uses only temporally available records to construct learner features and links alerts through attributions, evidence, and raw records.Future assessments and final outcomes are excluded from prediction-time data.Source: §3.1 · p. 5
3.2 Overview of the Incremental Learner Digital Twin
- The twin combines a growing observation layer with checkpoint-versioned analytic artifacts, preserving earlier features, attributions, and alerts for historical queries.The same graph supports feature extraction and traceability at each checkpoint.Source: §3.2 · pp. 6–7
3.3 Ontology and Evidence Path
- The ontology represents platform observations and analytic outputs, linking features to supporting evidence and alerts to their prediction runs, attributions, and selected evidence.Evidence items carry descriptions, strength, temporal scope, and links to raw records.Source: §3.3 · p. 7, Figure 2 · p. 8
- An alert’s evidence path makes its explanation queryable, verifiable, and traceable through standard graph queries.Checkpoint updates add newly available records, validate graph structure with SHACL, extract and materialize features, score learners, and write back alerts and evidence links.Source: §3.3 · p. 8, §3.4 · pp. 8–9
3.4 Synchronization, Feature Extraction, and Alert Grounding
- The checkpoint update process replays newly available records, validates required graph structure, and writes prediction features, alerts, attributions, and evidence links back into the twin.The output graph at one checkpoint becomes the input for the next.Source: §3.4 · pp. 8–9
4 Experimentation
The evaluation tests whether an incrementally updated learner twin can support temporal early-warning prediction while preserving equivalent raw-table features and grounding alerts in traceable evidence. On OULAD, the results report predictive parity, incremental graph growth with zero SHACL violations, and evidence chains linked to raw records.
4.1 Dataset and Temporal Protocol
- The experiment uses OULAD and trains on 2013 presentations before testing on 2014 presentations of the same modules at Weeks 4, 8, and 12.Learners with no activity or assessment before the first checkpoint are retained because their emptiness generates absence evidence; six modules meet the filtered protocol.Source: §4.1 · pp. 9–10
4.2 Baselines and Evaluation Metrics
- Seven model families are paired across raw-table and KG-mediated features using the same feature definitions, with F1 and AUC reported at each checkpoint.The evaluation also tracks incremental construction, SHACL violations, evidence coverage, presented fidelity, and traceability.Source: §4.2 · p. 10
4.3 Results
- By Week 12, the graph averages 321,401 cumulative learning traces and 3.28M triples, with zero SHACL violations at all checkpoints.Newly integrated traces average 148,560 at Week 4, 106,204 at Week 8, and 66,638 at Week 12 across six test modules.Source: §4.3 · p. 11, Table 2 · p. 11
- KG-mediated features preserve raw-table predictive results for every model, checkpoint, and metric on OULAD, with performance improving from Week 4 to Week 12.The strongest model varies by checkpoint: TABTX at Week 4 (AUC 0.735), XGB at Week 8 (AUC 0.809), and ET at Week 12 (AUC 0.847).Source: §4.3 · pp. 11–12, Table 3 · p. 12
- Contextual Cov@3, Contextual Cov@5, and traceability are each 1.000 at every checkpoint for the fixed XGB-ldt alert generator.Direct coverage is lower because absence and inactivity features use contextual records, while Fid@3 decreases later as evidence selection broadens.Source: §4.3 · p. 12, Table 4 · p. 12
- Week 8 case explanations connect risk alerts to attributed features, checkpoint-bounded evidence items, and originating records, including a false positive whose later assessment was excluded.The false-positive learner submitted the assessment on day 57, after the day-56 graph checkpoint.Source: §4.3 · pp. 12–13
5 Discussion
The discussion reports predictive parity between KG-mediated and raw-table features, alongside checkpointed graph coherence and evidence chains linking alerts to source records. It limits the demonstrated Learner Digital Twin to historical replay at predefined checkpoints, not live operation.
5 Discussion
- Identical F1 and AUC values reflect shared feature definitions; evidence chains link alerts to records, while absence and inactivity rely on contextual records.Registrations and due-assessment metadata provide contextual support rather than direct activity rows.Source: §5 · p. 13
- Using the graph for both feature derivation and evidential support reduces inconsistency risk and makes explanation completeness verifiable through SHACL constraints and graph queries.The graph grows from Week 4 to Week 12 with no SHACL violations.Source: §5 · p. 13
- The twin simulates synchronization by replaying historical data at predefined checkpoints; live ingestion, out-of-order events, corrections, integration, and feedback loops were not evaluated.The demonstrated system is a checkpoint-indexed twin with recurrent synchronization and analytic history, not a fully deployed live educational twin.Source: §5 · p. 13
6 Conclusion
The conclusion presents an incremental knowledge-graph Learner Digital Twin that represents prediction and evidence entities in a shared semantic structure. On OULAD, graph-mediated features preserve raw-table predictive behavior across model families and checkpoints, while alerts link to records through queryable evidence chains.
6 Conclusion
- The twin represents observations, predictive features, attributions, alerts, evidence items, and raw records as graph entities, serving both feature extraction and evidence grounding.This routes prediction through a semantic representation.Source: §6 · pp. 13–14
- On OULAD, graph-mediated feature extraction preserves equivalent raw-table predictive behavior across model families and checkpoints, while alerts link attributions and evidence to raw records.The evidence chains are queryable; the conclusion characterizes the approach as transparent, traceable, and verifiable without reducing predictive performance.Source: §6 · p. 14
- Future work will extend the twin toward competency-aware learner states, test it on live platforms, and examine effects on instructor understanding, trust, and intervention decisions.Source: §6 · p. 14