Contrastive Embedding Techniques Uncover Learner-Item Relationships in Digital Education Platforms

Abstract

For better personalization and performance of learners in digital education systems, it is essential to analyze the interplay between the learning materials themselves, the learners, and the educational institution. Detecting semantic similarity between learning items and learners is hard for classical recommendation models and performance prediction models, and they are hard to implement adaptive learning. This paper introduces an innovative CLIEN model that uses contrastive representation learning to learn embeddings of the learner and the educational items together in one space. The proposed model incorporates interaction history of the learner, evaluation outcomes, temporal patterns of learning and item meta data into a dual-encoder and contrastive loss-based structure to learn the most similarity between two relevant (learner, item) and minimize the similarity between two non-relevant (learner, item) objects. An adaptive attention mechanism helps refine the learned representations based on important behavioral features. Benchmark educational datasets were used to conduct experiments on the proposed model through an 80-20 split of training and testing data. The proposed CLIEN model is compared with collaborative filtering, matrix factorization, Deep Knowledge Tracing (DKT), and neural recommendation model. Experimental evaluation on benchmark educational datasets demonstrates that the proposed CLIEN framework consistently outperforms established baseline methods, including Collaborative Filtering (CF), Matrix Factorization (MF), Deep Knowledge Tracing (DKT), Neural Collaborative Filtering (NCF), and GNN-based Recommendation models. The proposed framework achieves superior Accuracy, Precision, Recall, F1-score, AUC, and NDCG@10, confirming the effectiveness of contrastive learner-item embedding for personalized educational recommendation. The learned embedding space enables discovering the hidden relationships between the learners and the items, leading to the more accurate recommendation of items to the learners.