Existing multi-instance learning (MIL) methods have advanced in bag-level label prediction but lack accurate instance-level label inference, a key requirement for fine-grained interpretability and annotation. To address this and boost classifier performance, two improvements are critical: (1) generating high-confidence instance pseudo-labels to resolve the ambiguity of unlabeled instances in bags, and (2) building an instance-level training framework adaptive to complex MIL data distributions and correlations. Therefore, an enhanced instance-level multi-instance classifier via attention mechanism and self-paced learning EMI-AM-SL was proposed. Specifically, the attention mechanism was used to dynamically highlight key instances (distinguishing positive instances in positive bags and suppressing noise), quantify feature importance, and generate credible pseudo-labels by measuring instance contributions to bag labels. Inspired by the "from easy to difficult" human cognition, the self-paced learning strategy was used to adjust instance training priority based on pseudo-labels and feature complexity, enabling progressive learning from simple, reliable instances to complex ones for better capture of bag-level discriminative patterns. Moreover, attention-derived instance and feature weights enhance model interpretability. Experiments on multiple MIL datasets show EMI-AM-SL outperforms comparative methods in generalization and interpretability in most cases.