Individual Prediction Performance Consistency-oriented Knowledge Tracing Model

XES3G5M was collected from an online learning platform developed by TAL Education Group and represents a large-scale Chinese-language mathematics interaction dataset. All experiments followed the standard data splitting strategy of pyKT, with random 80%/20% training-to-test set partitioning.For each individual student, the Area Under the ROC Curve (AUC) was independently computed on the test set. From these individual metrics, four evaluation indicators were aggregated for each model: the population-level aggregated AUC (AUC_agg), the arithmetic mean of per-student AUC values (AUC_avg), the standard deviation of per-student AUC values (std_AUC), and the range of per-student AUC values (R_AUC, i.e., maximum minus minimum).

Every experiment was repeated five times with independent random seeds, and the reported values represent the mean plus-minus one standard deviation across these five runs.2. Data DescriptionThe dataset consists of four comma-separated value (CSV) files encoded in UTF-8. Each row corresponds to the evaluation results of a specific model under a given experimental condition.

All metric values are dimensionless scalars (AUC ranges from 0 to 1; parameter counts are given in millions, M). No missing values are present. Because all experiments were averaged over five random-seed repetitions, the reported values are subject to statistical fluctuations introduced by random initialization; the corresponding standard deviations are reported alongside the means using the plus-minus notation.2.1 overall_experimental_results.csv — Overall Experimental ResultsThis file contains the overall comparison results between the proposed ConKT model and 12 representative baseline models across three datasets, totaling 39 data rows (including the header row).Row labels: The Dataset column indicates the public benchmark dataset used (AS2009, NIPS34, or XES3G5M).

The Model column includes DKT, DKVMN, DKT+, DeepIRT, GKT, KQN, SAKT, SKVMN, AKT, SAINT, ATKT, HawkesKT, IEKT, DIMKT, simpleKT, stableKT, lefoKT, UKT, ConKT-L, and the proposed ConKT, spanning representative knowledge tracing methods based on recurrent neural networks, attention mechanisms, and graph neural networks.Column labels: AUC_agg(up) denotes the population-level aggregated AUC, with higher values indicating better performance.

AUC_avg(up) denotes the arithmetic mean of per-student AUC values, reflecting the model's average predictive capability at the individual student level; higher is better. std_AUC(down) denotes the standard deviation of per-student AUC values, reflecting the inter-individual dispersion of model predictive performance; lower is better. R_AUC(down) denotes the range (maximum minus minimum) of per-student AUC values, reflecting the extreme disparities in individual-level prediction; lower is better.2.2 sequence_length_analysis_nips34.csv — Sequence Length AnalysisThis file records the comparison results of three representative models—ConKT, IEKT, and DKT—on the NIPS34 dataset under three different fixed sequence-length settings (L = 100, 200, and 400), totaling 10 data rows (including the header row).

The column labels follow the same definitions as in Section 2.1. An additional L column indicates the truncated student interaction sequence length.2.3 ablation_study.csv — Ablation Study ResultsThis file documents the ablation study results for each component of ConKT, totaling 13 data rows (including the header row). Experiments were conducted on both AS2009 and NIPS34 datasets.Model-variant labels: w/o Interaction Branch denotes removal of the interaction branch module. w/o Mastery Branch denotes removal of the mastery branch module. w/o Dual-Channel Rep. denotes removal of the question–knowledge-concept dual-channel representation module. w/ Mamba denotes replacement of the GRU-based dual-branch module with Mamba. w/o IRT-Fusion denotes removal of the IRT-based information fusion module.

ConKT denotes the full model. Column labels follow the same definitions as in Section 2.1.2.4 model_capacity_comparison.csv — Model Capacity Comparison ResultsThis file contains the controlled-capacity comparison results, totaling 12 data rows (including the header row). Experiments were conducted on both AS2009 and NIPS34 datasets.

By adjusting embedding dimensions and other architectural factors, the parameter counts of all comparison models were kept in the same order of magnitude as ConKT (approximately 4.11M), in order to verify whether ConKT's performance gains stem from increased model capacity. An additional column, Parameters(M), denotes the total number of model parameters in millions. All other column labels follow the same definitions as in Section 2.1.3.

File Overviewoverall_experimental_results.csv contains 39 rows of overall comparison results across three datasets. sequence_length_analysis_nips34.csv contains 10 rows of individual predictive consistency analysis under varying sequence lengths. ablation_study.csv contains 13 rows of ablation study results for ConKT components. model_capacity_comparison.csv contains 12 rows of controlled model capacity comparison results.

The above CSV files can be opened and processed using Microsoft Excel, LibreOffice Calc, Python (pandas), or any text editor that supports UTF-8 encoding.