Heterologous interest point matching method and device based on graph neural network

Abstract

The invention relates to a heterogeneous interest point matching method and device based on a graph neural network. The method comprises the following steps: acquiring a first interest point set and a second interest point set in a target territorial scope, constructing a first interest point atlas according to the first interest point set, and constructing a second interest point atlas according to the second interest point set; screening out a plurality of preliminary interest point matching pairs from the first interest point set and the second interest point set, and labeling the plurality of preliminary interest point matching pairs to obtain a plurality of seed interest point matching pairs; performing iterative training on a graph neural network model according to the first point-of-interest graph, the second point-of-interest graph and the plurality of pairs of seed point-of-interest matching pairs to obtain a trained graph neural network model; and processing the first point-of-interest atlas and the second point-of-interest atlas through a trained model, and determining all point-of-interest matching pairs in the first point-of-interest set and the second point-of-interest set according to a processing result. According to the method, the matching accuracy and the matching speed of the heterogenous interest points can be improved.

Description

technical field [0001] The present application relates to the field of electronic maps, in particular to a graph neural network-based heterogeneous interest point matching method, device, computer equipment and storage medium. Background technique [0002] Point of Interest (POI for short), which generally includes information such as name, address, latitude and longitude, category, etc., is the most important content of network electronic maps and the foundation of Internet location services. Due to the different sources of point-of-interest data on the Internet, the collection and processing processes are different, resulting in certain differences in the spatial location, attribute information and richness of these data. Therefore, how to effectively eliminate the inconsistency between data , and organizing them into a set of data that is accurate in content and available to users has become a hot spot of current research. [0003] Interest point matching is the process ...

AI Technical Summary

Problems solved by technology

[0013] (1) Based on the unsupervised similarity score method, the matching effect is poor for scenes where two interest points actually have a matching relationship, but the texts are quite different

[0014] (2) Although the two interest points are very close in text, the data that does not actually have a matching relationship will cause a mismatch

[0015] (3) The threshold of the similarity score is not easy to set

[0017] (1) A lot of feature engineering work is required to construct features, and the process is cumbersome

[0018] (2) The model is shallow, the expression ability is limited, and the ceiling of text matching effect is low

[0019] (3) The matching process is to match one-to-one interest points, and then traverse all the data to match as a whole. The matching efficiency is low, and it cannot be directly matched at the overall data level.

[0020] (4) This method assumes that the interest points are independent of each other. However, there is a certain spatial position relationship between the actual interest points, so the relationship information between the interest points is not used for matching, and the use of information is less, and the effect is not good.

[0022] (1) Pre-trained deep models generally input plain text information, which is poorly compatible with non-text features

[0023] (2) The pre-trained deep model is the same as the traditional machine learning model. The matching process is to match one-to-one interest points, and then traverse all the data to match as a whole. The matching efficiency is low, and it cannot be directly matched at the overall data level.

[0024] (3) This method assumes that the interest points are independent of each other. However, there is a certain spatial position relationship between the actual interest points, so the relationship information between the interest points is not used for matching, and the use of information is less, and the effect is not good.

Images