Abstract
Learning representations from large-scale, heterogeneous graphs can present challenges related to memory constraints and the prevalence of diverse, unstructured, or missing attributes. A disclosed system may address these limitations by employing an inductive graph neural network that can be trained on sampled subgraphs to improve scalability. The method may utilize a per-attribute aggregation strategy, for instance, using separate, learnable aggregators for different data modalities such as text, numerical, or categorical features. To handle incomplete data, the system can substitute a learnable vector for each missing attribute type, which may allow the model to learn a representation for the concept of missingness. This process can generate low-dimensional node embeddings that are inductive, enabling their application to new or unseen data for tasks such as recommendation, fraud detection, and semantic search in complex graph environments.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Kuligin, Leonid and Jakubowski, Wiktor, "Per-Attribute Aggregation and Learnable Missing Attribute Modeling", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/9777