A Method of Identifying Causes of Prediction Errors to Accelerate MLOps
Keita Sakuma, Ryuta Matsuno, 義男 亀田 · 2023
MLOps, techniques for solving operational problems in machine learning, have attracted attention in recent years. In MLOps, understanding why mispredictions occur is essential to improving operational models. However, misprediction analysis is currently a time-consuming process performed manually by data scientists. In this paper, we propose a flowchart-structured analysis method (called AIEDF) that automatically identifies the causes of mispredictions during model operation. Thanks to the flowchart structure, AIEDF's analyses are comprehensive and explainable. In addition, AIEDF is flexible in implementation and can be model agnostic. Through experiments with synthetic and real data, we have demonstrated that AIEDF accurately identifies root causes and provides valuable insights for model improvement.