The rapid expansion of online education has provided students with a vast array of learning resources. However, this abundance has also led to knowledge overload and the challenge of selecting appropriate learning materials. To address this issue, exercise recommendation has become an essential strategy for supporting personalized learning. While traditional collaborative filtering and knowledge tracing (KT) methods offer viable solutions for personalized recommendations, they often fall short in fully leveraging the rich semantic information embedded within exercise texts. In this study, we introduce a feature-aligned knowledge tracing-based exercise recommendation method, called FAKT-ER, which integrates the semantic understanding power of large language models (LLMs) with deep knowledge tracing (DKT) techniques. By aligning semantic and collaborative data, FAKT-ER enables a more accurate representation of students’ knowledge states, leading to improved exercise recommendations. Additionally, we incorporate feature dimensionality reduction and summation to enhance the model’s computational efficiency and expressive capacity. Experimental results across three educational datasets demonstrate that FAKT-ER outperforms other methods in terms of recommendation diversity, novelty, and accuracy, significantly enhancing the students’ online learning experience.