In modern ecommerce platforms, product content information may have two origins: one is tree-structured taxonomy attributes, and the other is free-form folksonomy tags. This paper proposes a hybrid model to incorporate taxonomy and folksonomy information to enhance ecommerce recommendations. It first develops a tree matching algorithm to establish the overall similarity between items, where tag information is integrated for semantic analysis for taxonomy attributes. Next, it proposes a unique random walk model on a heterogeneous graph constructed by user nodes and item nodes and different types of relations — user–item preference and item–item similarity relations. The random walk model is designed to be effective to identify the nearest item nodes for a particular user node, which are seen as the best-fit items for recommendations. Empirical experiments demonstrate that the proposed model improves performance in terms of both recommendation coverage and accuracy, especially for sparse data.