A Hybrid Siamese Neural Network for Personalized Learning Pathways in STEM Education

Abstract

Personalized learning paths have become important in STEM education due to differences in learning capacities, interest in learning, and progressions of learners. Conventional methods of adapting and suggesting new content to learners rely on rule-based and basic machine learning approaches that fail to account for the complexities of the learner-learner connection and dynamic educational processes. This paper presents a Hybrid Siamese Neural Network (HSNN) for building intelligent and adaptive personalized learning paths in the science, technology, engineering, and mathematics (STEM) education domain. A Hybrid Siamese Neural Network combines Siamese similarity learning, attention-based feature fusion, and educational analysis to establish similar learning paths for learners while considering their individual attributes. The HSNN uses multi-dimensional information about learners' characteristics, which include learner performance, engagement, and progression. In particular, the paper describes the design and implementation of a Siamese neural network that can establish similarity relations between learners. Moreover, the paper proposes the use of attention mechanisms for focusing on relevant educational information and improving the accuracy of learner recommendations. Various types of metrics were adopted to assess the efficacy of the recommended system, including metrics related to classification performance (accuracy, precision, recall, and F1 score), as well as those associated with educational performance (educational effectiveness). Based on experimental studies, it was found that the developed HSNN model considerably outperformed other approaches to personalized learning that included collaborative filtering, content-based recommendation systems, and traditional neural networks. The performance was quite high because the accuracy was 94.2% while the F1 score was 93.2%. In addition, the recommendation performance was highly reliable since the proposed model allowed learners' engagement to be increased by 18% while dropping the dropout rate by 15%. This means that the proposed model was quite efficient in enhancing personalized STEM learning. In summary, the deep similarity learning and attention mechanisms are highly efficient to enhance personalized learning systems.