Imbalanced multiclass classification with active learning in strip rolling process

Abstract

In the strip rolling process, conventional supervised methods cannot effectively address data with an imbalanced number of normal and faulty instances. In this paper, based on a deep belief network, a resampling method is combined with active learning (AL) to address imbalanced multiclass problems. The support vector machine-based synthetic minority oversampling technique was adapted to enrich the training data, whereas the true data distribution and model generalization were changed. A new selection strategy of AL is proposed that forms a function using uncertainty and diversity. AL uses an optimizing set that has a similar distribution with the whole dataset to calculate the informativeness of instances to optimize the model. Based on this step, the model study instances approach decision boundaries to promote performance. The proposed method is validated by five UCI benchmark datasets and strip rolling data, and experiments show that it outperforms conventional methods in tackling imbalanced multiclass problems.