This is an initial example demonstration of reproducibility for the KDD 2024 working paper titled "Enhancing Knowledge Retention: A Fusion Transformer Model with Knowledge Graph Embeddings."
Knowledge learning and retention are perennial challenges for human beings. In this study, we develop a machine learning model to predict the success rate of learning each knowledge entity in an upcoming session based on the user’s learning history. We begin by learning the latent representation of each knowledge entity within a knowledge graph. This graph captures both the inherent intellectual relationships between different knowledge entities and the learning patterns of individuals across these entities. Leveraging this informative knowledge graph, we extract the latent representation of each knowledge entity, which is then used to represent the user’s learning history. Employing the Transformer model to process the user’s learning history, our goal is to predict the success rate of learning each knowledge entity in the upcoming session. This integration of two modules enables us to forecast learner recall probabilities with unprecedented precision. By evaluating the model on 23 million study logs from an English vocabulary learning platform, we demonstrate that our proposed model surpasses state-of-the-art benchmarks, achieving a 13.14% increase in knowledge retention prediction accuracy and a 33.3% rise in the AUC metric. These significant improvements not only establish a new benchmark in the field but also pave the way for enhancing digital learning platforms, promising a more personalized, efficient, and effective learning process.