<p>Factorization machines (FMs) are well-known general predictors that efficiently model second-order feature interactions through inner products, even under significant sparsity. However, a key limitation of FMs lies in their equal treatment of interactions between different field pairs. To address this issue, subsequent studies have proposed capturing interaction differences by learning field specific embedding vectors or incorporating additional trainable weights. While these methods are able to distinguish the importance of different field interactions, they introduce challenges such as increased computational costs and parameter explosion. Additionally, they often fail to leverage inherent information in the original data, resulting in the loss of valuable insights and reduced interpretability. In this paper, we propose a novel lightweight and interpretable model called the <i>field-enhancing factorization machine</i> (FeFM). Specifically, we introduce a mutual information-based method that captures interaction differences between field pairs by measuring their strengths and using pre-computed interaction strengths as weights. This approach enables us to model differences in interaction strengths between field pairs without introducing additional parameters, while also supporting feature interaction selection to some extent. Extensive experiments on five real-world datasets demonstrate the advantages of our proposed FeFM compared to state-of-the-art methods.</p>