Systems and methods for enhancing an artificial intelligence model via a multi-field embedding approach

Abstract

The present disclosure provides a framework for enhancing artificial intelligence models while reducing its memory requirements, for example, via a multi-field embedding approach. In one exemplary aspect, this framework may be applied to knowledge graphs to mitigate the effects of node degree extremity and to reduce embedding memory demand. For example, this framework may provide multiple graph fields in knowledge graphs that embed nodes in different dimensions to correct for skewness in the distribution of node degrees and decrease memory requirements when training KGE models. In some aspects, the framework may implement a user-specified strategy to split a knowledge graph into a number of distinct fields with similar connectivity patterns. KGEs may then be computed in a model-agnostic manner to allow each field to embed into a different dimensionality, and mathematical transformations may be used to convert embedding dimensions between fields when needed during algorithmic processing.