Multi-Domain Merging Adaptation for Container Rehandling Probability Prediction

Abstract

The application of deep learning techniques to predict the container rehandling probability can assist yards in generating container storage strategies to reduce the rehandling rate of container within the yard. However, since the container storage dataset collected from the yard suffers from multiple sources and data drift, it is difficult to guarantee that the container rehandling probability prediction model directly trained on container storage dataset performs consistently well on different container blocks. To address above issue, we take each container block as an independent data domain and propose a novel multi-domain merging adaptation (MDMA) approach, which includes the three stages of domain merging, domain adaptation and aggregation. In domain merging stage, we merge data domains into several domain cluster for limiting the number of the source cluster and boosting the knowledge transfer. In domain adaptation stage, the prediction model learns the domain-invariant representation for each pair of the source and target clusters by distribution matching manner. In aggregation stage, the domain- invariant features are adaptively aggregates together by feed-forward network. Extensive experiments are conducted on the container rehandling dataset, and the results demonstrate the performance and practical values of the proposed MDMA approach.