测绘学报 ›› 2023, Vol. 52 ›› Issue (2): 318-328.doi: 10.11947/j.AGCS.2023.20210451

• 地图学与地理信息 • 上一篇    下一篇

面向不平衡POI类别分布的电子地图多层次标签预测模型

禹文豪1,2, 魏铖1, 陈佳鑫2   

  1. 1. 中国地质大学(武汉)地理与信息工程学院, 湖北 武汉 430074;
    2. 中国地质大学(武汉)国家地理信息系统工程技术研究中心, 湖北 武汉 430074
  • 收稿日期:2021-08-09 修回日期:2022-05-09 发布日期:2023-03-07
  • 作者简介:禹文豪(1987-),男,教授,博士生导师,研究方向为地图综合和空间数据挖掘。E-mail:ywh_whu@126.com
  • 基金资助:
    国家自然科学基金(42071442);中国地质大学(武汉)中央高校基本科研专项资金(CUG170640)

Predicting the unbalanced labels of POIs on digital maps using hierarchical model

YU Wenhao1,2, WEI Cheng1, CHEN Jiaxin2   

  1. 1. School of Geography and Information Engineering, China University of Geosciences, Wuhan 430074, China;
    2. National Engineering Research for Geographic Information System, China University of Geosciences, Wuhan 430074, China
  • Received:2021-08-09 Revised:2022-05-09 Published:2023-03-07
  • Supported by:
    The National Natural Science Foundation of China (No. 42071442);Fundamental Research Funds for the Central Universities, China University of Geosciences (Wuhan) (No. CUG170640)

摘要: 兴趣点(POI)是电子地图、导航等应用关注的主要要素之一,其数据质量直接影响地理信息服务的智能化水平。鉴于OpenStreetMap (OSM)等众源地理信息数据的非专业收集特征,其POI数据标签常存在缺失、标记错误等质量问题,亟须对POI标签进行智能化推断和增强处理。常规神经网络模型直接从单一层次预测多类别数据,未考虑POI类别在数量上分布不平衡的问题,其预测标签倾向于包含较多数据的类别,学习算法难以泛化小规模样本规则。本文考虑到不同POI类别间的数据规模差异较大,提出基于多层次POI类别组织的神经网络预测方法,通过小样本类别的层次化聚合,建立POI类别树结构,在树结构的不同层次上实现数据规模相对平衡的类别划分,支持神经网络高精度的标签预测。试验表明,本文方法仅需利用POI基础位置信息与邻近关系,其预测精度高于传统方法。

关键词: POI标签, 深度学习, 神经网络, 多层次模型

Abstract: Point of interest (POI) is one of the main elements of electronic maps, navigation and other applications. Its data quality directly affects the level of intelligence of geographic information services. In view of the non-professional collection characteristics of data on public geographic information platforms such as OpenStreetMap (OSM), the POI data labels often have quality problems such as missing labels or incorrect labels. Thus, there is an urgent need for intelligent inference of POI labels. The conventional neural network model predicts multi-category data labels directly from a single level, which does not consider the problem of the uneven distribution of POI categories. The labels predicted by neural network tend to data categories which contain larger data volume, where the learning algorithm is difficult to generalize small-scale sample rules. This paper takes into account the massive gaps in the data scale between different POI categories, proposing a neural network prediction method based on multi-level POI category organization. Through the hierarchical aggregation of small sample categories, the structured POI category tree is established, achieving a relatively balanced category division of the data scales at different levels of the tree, which supports the high-precision prediction of labels. Experiments show that based only on the POI location information, the accuracy of this method is higher than those of the traditional methods.

Key words: POI label, deep learning, neural network, multi-level model

中图分类号: