Journal
SOFT COMPUTING
Volume 25, Issue 21, Pages 13425-13435Publisher
SPRINGER
DOI: 10.1007/s00500-021-06090-9
Keywords
Network representation learning; Heterogeneous information networks; Metapath; Metawalk
Categories
Funding
- Ministry of Science and Technology (MOST) of Taiwan [109-2636-E-006-017, 110-2221-E-006-001, 110-2221-E-006 -136MY3]
Ask authors/readers for more resources
Network representation learning (NRL) is effective in generating node embeddings. However, existing studies on heterogeneous link prediction suffer from drawbacks such as the need for templates, separate embedding learning, and low-quality embeddings in incomplete networks. This work proposes a template-free method, metawalk2vec, which allows random walkers to visit diverse nodes, leading to improved node embeddings. Experiments on social and adoption link predictions show that metawalk2vec outperforms template-based models and is more robust to network incompleteness.
Network representation learning (NRL) is effective in generating node embeddings. To predict heterogeneous links between different types of nodes, NRL is not robustly investigated yet. Though existing studies on random walk-based heterogeneous NRL are available, it suffers from three drawbacks: need to specify templates (e.g., metapaths), require separate embedding learning in predicting heterogeneous links, and opt to generate low-quality embeddings when networks are incomplete or sparse. This work proposes a novel template-free NRL method, metawalk2vec, to tackle these issues for heterogeneous link prediction. The idea is allowing the random walker to visit diverse types of nodes, instead of following the pre-defined templates. While template-based methods use common context patterns for NRL, nodes depicted by uncommon context types can make their embeddings better distinguish from each other. We conduct the experiments of social (user-user) and adoption (user-item) link predictions on Twitter and Douban datasets. The results exhibit our metawalk2vec can achieve similar and even better performance than template-based models. We also show our model is more robust to the network incompleteness.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available