<p>In existing knowledge graph-based e-commerce product recommendations, initial representations may be biased due to inaccurate multimodal fusion, and embeddings can be further distorted by long-path high-order semantics propagation and the noise it introduces. Additionally, the model’s corrective ability is often limited by suboptimal negative sampling strategies. To systematically address these challenges, we propose a novel e-commerce product recommendation method, the Knowledge Graph-Based Global–Local Attention Recommender and Reinforced Negative Sampler (KG-GLARNS). Firstly, an initial representation method for users and products is proposed with the construction of an e-commerce multimodal knowledge graph. It fuses textual and visual features by integrating the contrastive language-image pre-training model with cross-attention mechanism. Fine-grained sentiment analysis of user reviews is incorporated to identify user preferences. Secondly, a recommendation module is constructed with a global–local attention mechanism employed to refine embeddings by learning global preference representations and task-specific local representations. Attention mechanism is used to highlight strongly relevant path information in the knowledge graph and suppress noise from irrelevant information. Finally, a reinforcement learning-based negative sampling module is introduced, in which informative negative-signal paths are inferred to dynamically generate higher-quality negative samples, with a reward function that penalizes false negatives. Experimental results on three publicly available datasets demonstrate the superior performance of the proposed KG-GLARNS model and verify the effectiveness of multimodal data and key modules. Specifically, it achieves average improvements of 9.97% in Recall@K and 13.36% in NDCG@K over the state-of-the-art baselines.</p>