Learning user preferences is critical in understanding users and recommending items. Previous works utilized user profiles or public reviews for user preference modeling. However, utilizing user profiles might cause privacy issues, and there are scenarios where items do not come with user reviews. In contrast, item metadata and user historical behaviors are easier to obtain. Moreover, item metadata are tag-formed features in many real-world scenarios. We aim to focus on the described scenario and utilize finer-grained features of the item and user historical behaviors to predict user preference. We proposed a model for unstructured feature combinations, which can handle unknown items, and provide experiments to justify the rationality of our model.