In recent years, recommender systems based on knowledge graphs (KGs) have gained significant attention. By effectively leveraging the rich factual information within KGs and capturing the relationships between items, these methods are able to provide more accurate recommendations. However, KG-based recommender systems still face challenges including missing facts and limited coverage, which restrict the in-depth understanding of items. Given the rich open-world knowledge and strong common-sense reasoning capabilities of large language models (LLMs), in this paper, we propose a novel method named EKLRec to address the above challenges. Specifically, EKLRec employs LLMs to understand and analyze items from the perspectives of item profiling and user preferences, and construct a dual-view recommendation framework. We also devise a knowledge fusion network to enhance the interaction and fusion of item features from different semantic spaces. Furthermore, an adaptive weighting multi-level contrastive learning module is proposed to promote the alignment of representations by maximizing the mutual information between different views. Extensive experiments conducted on three real-world datasets demonstrate the superior performance of our method over the state-of-the-arts.