Ada2vec: Adaptive Representation Learning for Large-Scale Dynamic Heterogeneous Networks

Representation learning generates the embedding vector of an object based on its relationships with others in a network. The generated vectors are inputs to various downstream machine learning tasks, such as classification, clustering and similarity search. The research area has attracted great interest and effort in recent years. However, due to its complexity, few of the existing representation learning methods have been developed for dynamic heterogeneous networks. Comparing with a static homogeneous network (graph), which contains single-typed objects (nodes) and relationships (edges) and remains unchanged over time, a dynamic heterogeneous network contains multiple-typed objects and relationships and evolves with time. We develop a novel adaptive representation learning algorithm, named Ada2vec, to address the challenges of embedding learning in large-scale dynamic heterogeneous networks. The key challenge is how to efficiently and effectively handle the network dynamics and heterogeneity features simultaneously. Ada2vec employs a statistical bound to capture network dynamics, and metapath-guided random walks to capture network heterogeneity. To the best of our knowledge, Ada2vec is the first approach that leverages the Hoeffding bound for modeling changes in dynamic heterogeneous network embedding. Extensive experiments demonstrate that Ada2vec significantly outperforms the state-of-the-art benchmarks in terms of embedding learning accuracy and efficiency. Compared to other dynamic heterogeneous network representation learning models, the size of the experiment datasets that we use are orders of magnitude larger (millions instead of tens of thousands).