Answers are grounded in this paper and cite their sources.
The traditional approach to machine learning often models problems as static phenomena, ignoring that we are dealing with a dynamic, high-volume environment. This paper addresses this issue by introducing a novel hybrid model combining a fast linear model to do near real-time predictions with the adaptability of deep neural networks in a streaming setup. The proposed approach utilizes a sliding time window to convert a static model approach into a streaming problem, allowing us to preserve anonymized and compressed historical data in a series of linear models. We show how to better model temporal features of the stream that can be otherwise lost when using a linear model, providing a more realistic representation of a dynamic environment. We introduce a metamodel to predict the next linear model in the data stream. We apply this method to an online advertising environment with the goal of user response prediction. We show the ability of the method to generalize features between streams while preserving user privacy. Experimental evaluations conducted on the iPinYou dataset demonstrate the superiority of the proposed approach over traditional linear models. The results highlight the model’s ability to adapt to temporal changes and to learn cross-advertiser patterns, fostering information sharing and improving predictive performance, showcasing its efficacy in the dynamic landscape of online advertising.
No takes yet. Share an insight, caveat, or question.
Wiatr et al. (2025) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: