Stable graph structure learning method and device

Abstract

The invention discloses a stable graph structure learning method and device, and relates to the field of machine learning, and the method comprises the steps: obtaining a graph structure implied witha biased relation in each environment through a data-driven mapping method according to a plurality of given biased sampling environments; obtaining the probability that each point in the graph structure is associated with a symbiotic data sample according to the graph structure of each environment and a multi-environment data generation mechanism; initializing a target graph structure, randomly sampling a symbiotic data sample in a plurality of environments, setting conditions for the symbiotic data sample, respectively obtaining a generation probability in each biased environment and a generation probability under an implicit relationship of the target graph structure, and taking the generation probability of the target graph structure as a mean value of the generation probability of each environment; and carrying out loop iteration until a convergence condition is satisfied so as to obtain a stable graph structure. According to the method, estimation deviation of parameters in a graph structure is reduced, and stable performance is kept in a complex and changeable test environment.

Description

technical field [0001] The invention relates to the technical field of machine learning, in particular to a method and device for learning a stable graph structure. Background technique [0002] Graphs are an effective representation method for describing rich and general relationships (edges) between things (points). The relational patterns (also known as knowledge) contained in the graph structure are also commonly used to meet the needs of various tasks. The classic data-driven mapping method is very sensitive to the distribution of sampled data due to the sparsity of samples and the deviation of model assumptions. If there is a selective bias in the sampling environment of the data, the high-order, nonlinear relationship depicted by the graph structure will also be biased, which will greatly affect the generalization performance of the graph. [0003] In the new research results, it is mainly aimed at the challenges brought about by the selection bias and sparseness of ...

AI Technical Summary

Problems solved by technology

There are the following challenges in related technologies: Challenge 1: Traditional data-driven mapping methods are highly dependent on the independent and identical distribution assumption of data generation, but in real scenarios due to the unknowability of data sources, this assumption is not easy to satisfy

The present invention proposes a stable learning algorithm framework for this purpose, aiming to eliminate the deviation of a single environment by using multi-environmental information; Challenge 2: Relative to the scale of the graph, there is a problem of data sparsity

This sparsity and systematic bias in mapping methods can further amplify sample selection bias

Images