Modern recommender systems increasingly leverage multi-behavior data to alleviate data sparsity, yet remain challenged by interrelated issues of noise interference and behavior distribution imbalance. While traditional approaches typically address these issues in isolation, we argue that noise in random interactions leads to an imbalanced distribution of behaviors, and correspondingly, the imbalanced distribution of user behaviors causes biases in the recommendation. To address this interrelation, we propose a unified model DiffHGCN. DiffHGCN proposes an asymmetric embedding learning approach for users and items in multi-behavior recommendation. User embeddings undergo behavior-specific diffusion denoising guided by target interaction frequencies to filter noise while preserving true preferences. Item embeddings employ lightweight global diffusion combined with behavior-specific propagation for computational efficiency. The framework dynamically adjusts behavior weights through self-attention mechanisms based on semantic relevance and frequency patterns. Experiments on three real-world datasets demonstrate that our model significantly outperforms baseline models.