Incomplete Multi-view Sequential Representation Learning with Cascaded Cross-view Evolution Attention

Multi-view sequential data are widely used in the real world, while representation learning can integrate heterogeneous data from different views to improve learning performance which makes significant progress in recent years. However, most existing algorithms assume that sequential data are complete, which frequently contradicts reality. Due to limitations in data collection and sensor reliability, multi-view sequence data often suffer from missing issues, which makes representation learning more challenging. To address the above problem, this paper proposes an incomplete multi-view representation learning algorithm with cascaded cross-view evolution attention network (named IMvCCEA), which aims to effectively integrate unique and shared information from different views and achieve robust representation learning of multi-view sequential data. Specifically, we first employ the dominant view-guided completion strategy, which utilizes semantically rich views as anchors to guide the fusion process and overcome the difficulty of feature fusion caused by missing data. Subsequently, we design a cascaded cross-view evolutionary attention module that gradually refines the dominant representation by fusing complementary information from auxiliary views to address the feature complementarity in incomplete views. Meanwhile, we introduce an adaptive dynamic integration mechanism, which adaptively balances the original and evolved feature representations based on model confidence to maintain consistency. We conduct extensive comparative experiments on various real-world datasets and comparison algorithms to verify the effectiveness of the proposed algorithm. The experimental results demonstrate that the proposed algorithm outperforms other comparison algorithms under different experimental settings, which proves its superiority. Further in-depth analysis shows that each key module in the framework has played an important role in improving performance, which verifies the rationality and practicality of the overall framework.