Lime: Low-Cost and Incremental Learning for Dynamic Heterogeneous Information Networks

Understanding the interconnected relationships of large-scale information networks like social, scholar and Internet Things is vital for tasks recommendation fraud detection. The vast majority real-world are inherently heterogeneous dynamic, containing many different types nodes edges can change drastically over time. dynamicity heterogeneity make it extremely challenging to reason about network structure. Unfortunately, existing approaches inadequate in modeling real-life dynamical as they either have strong assumption a given stochastic process or fail capture structure, all require extensive computational resources. We introduce Lime , better approach dynamic networks. designed extract high-quality representation with significantly lower memory resources time state-of-the-arts. Unlike prior work that uses vector encode each node, we exploit semantic among multiple similar semantics shared vectors. By using fewer node vectors, our reduces required space encoding To effectively trade sharing reduced footprint, employ recursive neural (RsNN) carefully optimization strategies explore novel cuboid space. then go further by showing, first time, how an effective incremental learning be developed – help RsNN, set techniques allow framework quickly efficiently adapt constantly evolving network. evaluate applying three representative network-based tasks, classification, clustering anomaly detection, performing on datasets. compare against eleven state-of-the-art representation. Our experiments demonstrate not only footprint 80 percent processing 2x when but also delivers comparable performance downstream tasks. show method boost up 20x without compromising quality learned