Network embedding learns low-dimensional representations of nodes with the goal of preserving the original network structure. However, most existing embedding methods lack the ability to handle event data, which are ubiquitous in the real world, due to the following three challenges: (1) participating objects in an event are often of different types, which limit the feasibility of using homogeneous network embedding methods; (2) relations among nodes in each event are much more complicated, i.e., more than two objects are involved in one event, thus it is far from enough to only preserve pairwise proximity; (3) there may exist relevance among different events, which has effects on the representations. In this paper, we model event data as a heterogeneous hypergraph, where participating objects in one event are represented as a hyperedge, and propose a novel embedding framework, namely event2vec, for learning effective representations of objects by preserving both the intra-event proximity and inter-event proximity. Extensive experiments on large-scale real-world datasets demonstrate that the representations learned by event2vec can outperform state-of-the-art methods.

View PDF No PDF? Try Arxiv!