This paper studies the Change Point Detection (CPD) problem, which detects abrupt changes in distributions in data streams. CPD for high-dimensional data remains challenging due to complex data patterns and violations of common assumptions. Existing deep learning-based CPD models fail to achieve perfect performance, and ensemble techniques offer a more robust solution. In this paper, we study deep CPD model ensembles and demonstrate that standard predictive aggregation techniques, such as averaging, are suboptimal. As an alternative, we propose WWAggr, a novel ensemble aggregation method based on the Wasserstein distance. This method effectively applies to various deep CPD model ensembles and addresses the problem of selecting a decision threshold for CPD.
