Bi-CLKT: Bi-Graph Contrastive Learning based Knowledge Tracing

Xiangyu Song, Jianxin Li, Qi Lei, Wei Zhao, Yunliang Chen, Ajmal Mian

Abstract

Knowledge Tracing (KT) is a critical task in intelligent tutoring systems, aimed at modeling a student's mastery of knowledge concepts over time. Traditional deep learning approaches often struggle with granularity mismatches between specific exercises and broader concepts, as well as the inherent sparsity of student-exercise interaction data. This paper introduces Bi-CLKT, a novel bi-graph contrastive learning framework. By constructing an exercise influence subgraph and employing a two-layer self-supervised learning scheme (node-level and graph-level), Bi-CLKT effectively captures both local exercise nuances and global concept relationships. Experimental results on four benchmark datasets demonstrate that Bi-CLKT significantly outperforms state-of-the-art models in predictive accuracy.

Introduction

Knowledge Tracing models a student's evolving knowledge state to predict their future performance. While Deep Knowledge Tracing (DKT) and Graph Neural Networks (GNNs) have advanced the field, they face three primary challenges:

  1. Granularity Mismatch: Models often fail to balance high-level semantic information with individual node details.
  2. Structural Complexity: Capturing the intricate spatial associations between different exercises remains difficult.
  3. Entity Isolation: Most models treat concepts (e.g., "Addition") and exercises (e.g., "5+7") as separate entities, ignoring that exercises within the same concept can vary significantly in difficulty and semantic nuance.

Bi-CLKT addresses these issues by integrating both "exercise-to-exercise" (E2E) and "concept-to-concept" (C2C) relationships into a unified contrastive learning framework.

Methodology

The Bi-Graph Structure

The core of Bi-CLKT is the transition from a simple exercise-to-concept mapping to a more complex exercise influence subgraph ($G_e$). In this graph, edges represent the correlation between exercises based on student performance patterns (e.g., co-occurrence of correct answers).

The transition from (a) basic exercise-to-concept mappings to (b) the exercise influence subgraph G_e, where exercises are connected based on mutual influence.

Two-Layer Contrastive Learning

The framework utilizes a dual-level self-supervised learning approach to extract robust representations:

  • Node-level (Local): Employs GCNs on subgraphs to learn discriminative representations of individual exercises (E2E).
  • Graph-level (Global): Learns representations of broader concepts (C2C) by treating subgraphs as single entities.

To facilitate contrastive learning, the model uses Adaptive Augmentation. It employs centrality measures like PageRank to prune unimportant edges while preserving critical structural nodes. This ensures the model learns the "core identity" of the graph even under noise.

The Bi-CLKT architecture, illustrating the flow from the influence subgraph through dual-level augmentation and GCN encoders to the final prediction layer.

Prediction Layer

The learned embeddings from both levels are concatenated and fed into a final prediction layer. The authors evaluated two variants: Bi-CLKT-R (utilizing an RNN for temporal modeling) and Bi-CLKT-M (utilizing a Memory-augmented neural network).

Experimental Evaluation

Datasets

The model was benchmarked against four real-world datasets: ASSISTment 2009, ASSISTment 2015, ASSISTment Challenge, and STATICS 2011. These datasets vary in size and concept density, providing a comprehensive test of the model's robustness.

Key Results

Bi-CLKT consistently outperformed all baselines, including high-performance Transformer-based models like SAINT+.

  • Predictive Accuracy: On the ASSISTment 2009 and STATICS 2011 datasets, Bi-CLKT achieved an AUC of approximately 0.86, a significant improvement (at least 5%) over traditional methods like DKT or DKVMN.
  • Ablation Studies: Results confirmed that concatenating E2E and C2C embeddings yields superior performance compared to using either in isolation, validating the bi-graph approach.
  • Temporal Modeling: The Bi-CLKT-R variant generally performed best, suggesting that the combination of high-quality graph embeddings with temporal RNN modeling is particularly effective for KT.

Conclusion

Bi-CLKT successfully bridges the gap between exercises and concepts by leveraging bi-graph contrastive learning. By recognizing that exercises are not merely proxies for concepts but have their own unique influence patterns, the model achieves state-of-the-art results. Future work will focus on improving the computational efficiency of influence graph construction and enhancing the interpretability of the learned representations for educational stakeholders.

Summary updated

BibTeX
@article{song2022bi,
  author = {Xiangyu Song and Jianxin Li and Qi Lei and Wei Zhao and Yunliang Chen and Ajmal Mian},
  title = {{Bi-CLKT: Bi-Graph Contrastive Learning based Knowledge Tracing}},
  journal = {Knowledge-Based Systems},
  year = {2022},
  doi = {10.1016/j.knosys.2022.108274},
}
Figure