The patent citation network is a complex and dynamic system that reflects the diffusion of knowledge and innovation across different fields of technology. With this work, we aim to analyze such citation networks by developing a novel approach that leverages Relational Event Models (REMs) and Machine learning concepts. Overcoming the main limitations of REMs on analyzing large sparse networks, we propose a Deep Relational Event Additive Model (DREAM) that models the relationships between cited and citing patents as events that occur over time, capturing the dynamic nature of the patent citation network. Each predictor in the generative model is assumed to have a non-linear behavior, which has been modeled through a B-spline approach that allowed us to capture such smooth effects. By estimating the model through a stochastic gradient descent approach, we were able to efficiently estimate the parameters of the DREAM and identify the key factors that drive the network dynamics. Additionally, our spline approach allowed us to capture complex relationships between predictors through elaborate interaction effects, leading to a more accurate and comprehensive interpretation of the underlying mechanisms of the patent citation network. Our analysis revealed several interesting insights, such as the identification of time windows in which citations are more likely to happen and the relevancy of the increasing number of citations received per patent. Overall, our results demonstrate the potential of the DREAM in capturing complex dynamics that arise in a large sparse network, maintaining the features and the interpretability for which REMs are mostly famous.

翻译：专利引用网络是一个复杂且动态的系统，反映了知识和技术创新在不同技术领域的扩散。本研究旨在通过开发一种融合关系事件模型（REMs）与机器学习概念的新方法，来解析此类引用网络。为克服REMs在分析大型稀疏网络时的主要局限性，我们提出了一种深度关系事件加性模型（DREAM），将引用专利与被引用专利之间的关系建模为随时间发生的事件，从而捕捉专利引用网络的动态特性。生成模型中的每个预测变量被假定为具有非线性行为，我们通过B样条方法建模这些平滑效应。通过采用随机梯度下降方法估计模型，我们能够高效估计DREAM的参数，并识别驱动网络动态的关键因素。此外，我们的样条方法通过复杂的交互效应捕捉预测变量间的复杂关系，从而更准确、更全面地解读专利引用网络的潜在机制。分析揭示了一些有趣的现象，例如识别出引用更可能发生的时间窗口，以及每项专利接收的引用数量增加的相关性。总体而言，我们的结果证明了DREAM在捕捉大型稀疏网络中复杂动态方面的潜力，同时保留了REMs最广为人知的特性和可解释性。