The cold-start problem remains a critical challenge in recommendation systems, particularly in cross-domain scenarios with Domain-Level Zero-Shot Recommendation (DZSR) tasks. Existing methods for DZSR fail to distinguish irrelevant triples in the knowledge graph (KG) and insufficient in capture users’ nonlinear features interests. To address these problems, we propose a Cross-domain Item Knowledge Graph Embedding (CIKGE) framework for DZSR. The core idea of our method is to construct a Cross-domain Item Knowledge Graph (CIKG), design a Cross-domain Item Embedding (CIE) to filter the irrelevant triples, and design a Cross-domain User Embedding (CUE) to capturing the nonlinear features of user interest. Specifically, we first select relevant cross-domain paths from the CIKG, constructing a two-level feature matrix to embed the relation paths between the source and target domains, thereby generating cross-domain item embeddings. Additionally, to address the complexity of user behavior, we perform clustering analysis on user interaction items using an Adaptive Growing Self-Organizing Map (AGSOM) to extract nonlinear interaction features, which are integrated into cross-domain user embeddings. Extensive experiments on real world datasets demonstrate that our CIKGE framework achieves superior performance compared to state-of-the-art baselines.