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

Abstract: The goal of Knowledge Tracing (KT) is to estimate how well students have mastered a concept based on their historical learning of related exercises. The benefit of knowledge tracing is that students' learning plans can be better organised and adjusted, and interventions can be made when necessary. With the recent rise of deep learning, Deep Knowledge Tracing (DKT) has utilised Recurrent Neural Networks (RNNs) to accomplish this task with some success. Other works have attempted to introduce Graph Neural Networks (GNNs) and redefine the task accordingly to achieve significant improvements. However, these efforts suffer from at least one of the following drawbacks: 1) they pay too much attention to details of the nodes rather than to high-level semantic information; 2) they struggle to effectively establish spatial associations and complex structures of the nodes; and 3) they represent either concepts or exercises only, without integrating them. Inspired by recent advances in self-supervised learning, we propose a Bi-Graph Contrastive Learning based Knowledge Tracing (Bi-CLKT) to address these limitations. Specifically, we design a two-layer contrastive learning scheme based on an "exercise-to-exercise" (E2E) relational subgraph. It involves node-level contrastive learning of subgraphs to obtain discriminative representations of exercises, and graph-level contrastive learning to obtain discriminative representations of concepts. Moreover, we designed a joint contrastive loss to obtain better representations and hence better prediction performance. Also, we explored two different variants, using RNN and memory-augmented neural networks as the prediction layer for comparison to obtain better representations of exercises and concepts respectively. Extensive experiments on four real-world datasets show that the proposed Bi-CLKT and its variants outperform other baseline models.

• The paper introduces Bi-CLKT, a novel knowledge tracing model utilizing bi-graph contrastive learning on both node-level and graph-level structures for comprehensive representation of educational interactions.
• Bi-CLKT consistently outperforms baseline models across four real-world datasets, demonstrating around 5% improvement in AUC and ACC by effectively integrating graph-based contrastive learning.
• This research highlights the effectiveness of multi-layered learning architectures in knowledge tracing, paving the way for more precise AI-assisted educational systems and broader applications of graph-based methods.

A Critical Analysis of Bi-CLKT: Bi-Graph Contrastive Learning based Knowledge Tracing

The paper "Bi-CLKT: Bi-Graph Contrastive Learning based Knowledge Tracing" introduces a novel approach in the domain of knowledge tracing (KT), leveraging graph-based contrastive learning to enhance prediction accuracy of student mastery over different concepts. In alignment with the growing shift towards deep learning methodologies, this study tackles existing limitations found in prior KT models that either excessively focus on node-level details or fail to integrate concept and exercise representation sufficiently. Bi-CLKT distinguishes itself through the use of Bi-Graph Contrastive Learning (CGN) applied on both node-level (local) and graph-level (global) structures, offering a more comprehensive representation of educational interactions.

Key Contributions and Methodological Details

The authors explore well-known issues associated with traditional KT models, specifically the inadequacy in representing exercises and concepts comprehensively. To address these, they propose a self-supervised learning framework that applies contrastive learning strategies to construct differentiated embeddings. This approach retains high semantic information by adopting what is referred to as "exercise-to-exercise" (E2E) and "concept-to-concept" (C2C) relationships. By doing so, Bi-CLKT effectively learns distinct and accurate patterns that enhance prediction performance.

Central to the Bi-CLKT strategy is the formation of exercise influence subgraphs derived from student transition patterns between exercises, hypothesizing that students solving different exercises correctly indicate implicit exercise similarity or relation. The use of node-level and graph-level GCNs aids in extracting these relational features, producing embeddings that encapsulate higher-order information rather than oversimplifying or isolating node details. This aspect of their methodology sets a precedent for future applications of graph-based strategies within KT tasks.

Experimental Results and Comparative Analysis

The evaluation employs four real-world datasets: ASSISTment 2009, ASSISTment 2015, ASSISTment Challenge, and STATICS 2011. Bi-CLKT consistently surpasses baseline models such as BKT, DKT, DKVMN, SAKT, and others in terms of both Area Under the Curve (AUC) and Accuracy (ACC). Notably, Bi-CLKT achieves a 5% improvement in these metrics, affirming the advantages of incorporating graph-based contrastive learning into KT methodologies.

Implications and Looking Ahead

The implications of this study are twofold. Practically, it signifies a shift towards more integrative, precise modeling in educational systems designed to trace student knowledge. Theoretically, it paves the way for more refined graph-based learning methods potentially applicable to a broader array of AI tasks beyond KT. This research contributes significantly to existing literature by synthesizing theories from graph neural networks, contrastive learning, and deep learning, thereby reinforcing the argument for multi-layered learning architectures.

Nonetheless, the practical application of Bi-CLKT warrants further exploration, particularly regarding data augmentation and graph configuration strategies that might enhance or detriment their approach's applicability in diverse educational scenarios. Future work could investigate alternative embedding techniques that further optimize the capture of semantic relationships within KT tasks.

In summary, Bi-CLKT presents a robust framework that addresses critical gaps in knowledge tracing models, heralding a promising future for AI-assisted education systems. Its contribution to the field is evident not only in its remarkable performance improvements but in its strategic methodological innovations that integrate self-supervised learning principles into traditional KT problems.