Deep learning for tabular data advance in extracting valuable information from column feature interactions. However, tabular data consists of various column feature types such as numerical, categorical and binary, each with distinct semantics, ranges and distributions. Existing methods fail to distinguish between the heterogeneity of features and their interactions, leading to irreversible information loss. To address this, we propose the Heterogeneous Feature Interaction Network (HFIN). HFIN represents column features as different types of nodes in a relational graph and enhances heterogeneous information in feature interactions through relational message passing. To preserve key information from heterogeneous features, we embed numerical, categorical, and binary features separately and integrate sequential information to improve numerical feature embeddings. We use mutual information and symmetric matrix factorization to estimate a global feature interaction graph, capturing both semantic and topological structures of feature interactions. Additionally, we introduce a relational attention-based message passing mechanism, which dynamically adjusts edge weights to capture critical information, further strengthening the expression of heterogeneous interactions. Experimental results on six datasets show that HFIN outperforms most DNN methods in classification and regression tasks, while providing better interpretability.