Online Learning for Wild Streaming Data with Delayed Feedback

Wild streaming data (i.e., streaming data with concept drift, missing values, and heterogeneous features) are often encountered in practical deployments. Online learning is an effective approach for handling wild streaming data by learning from incoming data streams only once. Existing methods typically assume that ground-truth labels are available without delay, enabling continuous updates of online learning. However, an often overlooked but critical issue is delayed feedback caused by the slow and expensive labeling process, which prevents timely updates of online learning. This paper investigates a new problem of Online Learning for Wild streaming data with Delayed Feedback (OLWDF). To this end, we propose an OLWDF algorithm by extracting geometric information from wild streaming data to mitigate the negative impact of delayed feedback. First, we employ an online Gaussian copula to project wild streaming data onto a unified and aligned latent metric space. Second, we construct a dynamic geometric graph for transforming topological evolution into soft pseudo-labels via an uncertainty-aware evidence calibration mechanism. Finally, we design a decoupled dual-track learning architecture, combining a label-driven persistent learner for stability and a geometry-guided transient learner for low-risk prospective adaptation. Extensive experiments on ten real-world datasets and four synthetic datasets demonstrate that our algorithm outperforms competing methods. The source code of this paper is available at the link: https://github.com/YulinWang-pro/OLWDF.