Abstract: In this paper, we propose a local outlier detection algorithm based on multi-instance learning (LOF-MIL). In our approach, polysemous objects are abstracted to a multi-instance using an MIL framework, then the MIL-LOF calculates the comprehensive outlier factor and detects outliers by adopting degradation strategies and making weight adjustments. We compared our approach with the classic local outlier detection algorithm and its optimization algorithm on both public and real data sets. Experimental results show that our method achieves better accuracy, comprehesiveness, and efficiency.