Searching to extrapolate embedding for out-of-graph node representation learning.

Abstract

Out-of-graph node representation learning aims at learning about newly arrived nodes for a dynamic graph. It has wide applications ranging from community detection, recommendation system to malware detection. Although existing methods can be adapted for out-of-graph node representation learning, real-world challenges such as fixed in-graph node embedding and data diversity essentially limit the performance of these methods. Here, we formulate the problem as a neural architecture search problem, and propose searching to extrapolate embedding (S2E), a solution that extrapolates embedding for out-of-graph nodes according to their neighbor node embeddings. Firstly, we propose an embedding extrapolating framework containing multiple transition modules and an aggregation module to handle fixed in-graph node embedding for embedding extrapolation. To deal with data diversity, we propose searching extrapolating architecture, where we employ objective transformation to handle non-differentiable evaluation metric and make neural architecture search procedure more efficient. In experiments, we show that S2E achieves outstanding performance in real-world datasets. We further conduct experiments on the proposed search space and search algorithm to verify the effectiveness of our design in S2E.