Traditional Conversational Recommendation Systems (CRS) typically assume that users have explicit preferences during interactions, overlooking the fact that user interests evolve over time and across different contexts. Additionally, users generally select only one preferred attribute per interaction, and rejected items are permanently removed from the candidate set, which may eventually lead to a lack of viable recommendations. To address these issues, we proposes a more practical conversational recommendation scenario—Knowledge-Enhanced Multi-round Conversational Recommendation (KEMCR), where rejected candidates are not permanently removed based on a single instance of feedback. Under this scenario, we propose a novel method, Knowledge-enhanced Multi-view Graph Contrastive Learning for Conversational Recommendation (KMGCL), which constructs more precise user representations from three perspectives: the current conversation view, the historical global view, and the similar user view. We employ graph convolutional networks to extract diverse and fine-grained features while integrating multi-view preferences and optimizing node embeddings through contrastive learning. Additionally, an attention mechanism is introduced to iteratively capture users’ multiple interests, and a noisy strategy network is adopted to learn and refine conversational strategies, enabling better handling of uncertainties in user interactions. Experimental results demonstrate that KMGCL significantly outperforms existing state-of-the-art conversational recommendation methods on benchmark datasets.