Feature selection for quality prediction under distribution shift

Distribution shifts due to the system and the external reasons in industrial processes are very common, which devastates the predictions of the linear model easily. This paper provides a practical solution for quality prediction under the circumstance of constant data shifts. Through analyzing a real-world petroleum plant data set, we show that via keeping the linear model as simple and relevant to the task as possible, the impacts of the data shift on the linear model can be diminished, under the assumption that the relationships between the target variable and the explanatory variables maintain in a similar state. In particular, we propose a pragmatic procedure for feature selection, including dealing with redundant features, considering interactions and enlarging the feature space. Extensive experiments demonstrate the effectiveness of this method, and comprehensive analysis of the results confirms and supports this finding.