Independent edge evaluation for directed acyclic graphs in probabilistic sequence modeling

This repository contains the official replication package and data artifacts for the Step-Level Diagnostic Engine (SLDE) framework and the Independent Edge Evaluation (IEE) methodology, ensuring complete reproducibility of the empirical results presented in the manuscript. The dataset consists of three primary components structured to support process-oriented assessment space evaluation. The first component is the ASSISTments 2009-2010 Dataset (stored in skill_builder_data.csv). This file contains the benchmark, publicly available raw sequential tracking data from the ASSISTments 2009-2010 skill-builder corpus. To isolate multi-step cognitive structures, the replication script applies strict preprocessing rules directly to this raw file. For main tasks (original=1), records with null skill names or non-binary correctness are discarded. For sub-step scaffold rows (original=0), records are retained and programmatically inherit latent dependency labels from the parent main task via the assistment_id. The final execution filters a restricted cohort of exactly M = 669 unique learners, structured to enforce zero temporal data leakage during downstream sequence modeling. The second component is the Prerequisite Graph Topology (stored in assistments_prereq_edges.sql). This file provides the explicit graph-theoretic formalization of the assessment space used to evaluate cross-skill transfer. Parsed directly via regular expressions within the scripts, it contains a mapping of 110 unique Knowledge Components (KCs), directed prerequisite edges mapping cognitive transitions, and a set of 857 pre-enumerated valid pedagogical paths ordered longest-first, acting as the ground-truth Directed Acyclic Graph (DAG) templates. The third component is the Complete Source Code Package (stored in IEE_DAG_Final_Submit.ipynb). This comprehensive Python replication notebook contains the complete implementation pipeline. It includes environment setup pinning sympy==1.13.1 to resolve native compatibility conflicts with PyTorch 2.x optimizers. It provides the exact mathematical architecture for the proposed IEE-BKT model (implementing localized evidence vectors and K-step sequential Bayesian updates) along with re-implemented baselines including Standard BKT, LSTM-based Deep Knowledge Tracing (DKT), and Self-Attentive Knowledge Tracing (SAKT). Finally, it executes a strict student-level 10-fold cross-validation routine to control for intra-student tracking bias. Researchers can utilize these files to fully replicate the evaluation metrics, including paired t-test distributions, RMSE improvements, and visual ROC/AUC patterns where IEE-BKT secures a baseline performance of AUC = 0.5514.本仓库包含面向步骤级诊断引擎（Step-Level Diagnostic Engine, SLDE）框架与独立边缘评估（Independent Edge Evaluation, IEE）方法的官方复现套件与数据制品，可完整复现论文中呈现的实证研究结果。本数据集包含三个核心组件，旨在支撑面向过程的评估空间评估。 第一组件为ASSISTments 2009-2010数据集（存储于skill_builder_data.csv）。该文件包含ASSISTments 2009-2010技能构建语料库中公开可用的基准原始序列追踪数据。为隔离多步认知结构，复现脚本会直接对该原始文件应用严格的预处理规则：针对主任务（原始标记为1），将丢弃技能名称为空或正确性标签非二元的记录；针对子步骤支架行（原始标记为0），则保留记录并通过assistment_id从父主任务中程序化继承潜在依赖标签。最终执行时会筛选出恰好包含M=669名唯一学习者的受限队列，以确保下游序列建模过程中无时间数据泄露。 第二组件为先决条件图拓扑结构（存储于assistments_prereq_edges.sql）。该文件提供了用于评估跨技能迁移的评估空间的显式图论形式化表达。通过脚本内的正则表达式直接解析，其包含110个唯一知识组件（Knowledge Components, KCs）的映射关系、表征认知迁移的有向先决边，以及857条预先枚举的、按长度降序排列的有效教学路径，作为基准有向无环图（Directed Acyclic Graph, DAG）模板。 第三组件为完整源代码套件（存储于IEE_DAG_Final_Submit.ipynb）。该综合性Python复现笔记本包含完整的实现流程：其中固定了sympy==1.13.1的环境配置，以解决与PyTorch 2.x优化器的原生兼容性冲突；提供了所提出的IEE-BKT模型的精确数学架构（实现了局域化证据向量与K步序列贝叶斯更新），同时复现了包括标准贝叶斯知识追踪（Bayesian Knowledge Tracing, BKT）、基于长短期记忆网络（Long Short-Term Memory, LSTM）的深度知识追踪（DKT）以及自注意力知识追踪（Self-Attentive Knowledge Tracing, SAKT）在内的基准模型；最后执行了严格的学习者级10折交叉验证流程，以控制学习者内部追踪偏差。 研究人员可利用这些文件完整复现各项评估指标，包括配对t检验分布、均方根误差（Root Mean Square Error, RMSE）提升效果，以及IEE-BKT达到AUC=0.5514基准性能的可视化受试者工作特征曲线（Receiver Operating Characteristic, ROC）与曲线下面积（Area Under Curve, AUC）模式。