Quality Prediction of Brewing Koji Based on Multitask Learning

Koji making is an important craft with critical significance for assessing the quality of liquor. However, under the same physicochemical environment, due to the difference in koji reproductive habits, different levels of koji have different content scores distribution, so, there are great challenges in predicting multiple scores of koji accurately. To address this issue, we proposed a multi-task learning approach by leveraging a shared bottom-level representation layer to extract shared features from the data and joint training on multiple koji rating tasks to learn the relationships between different levels of koji. We conduct experimental evaluations on a real-world koji making dataset. The results demonstrate that the proposed method exceeds other benchmark learning methods in predicting koji scores and capturing the relationships between different levels of koji, thereby enhancing the accuracy and robustness of predictions.