Automatic Recommendations of Categories for Geospatial Entities

Giorgos Giannopoulos, Nikos Karagiannakis, Dimitrios Skoutas, Spiros Athanasiou

Abstract

The OSMRec framework addresses the challenge of manual tagging in OpenStreetMap (OSM) by providing automated category recommendations. By leveraging a combination of geometric features and textual metadata, the system assists contributors in selecting appropriate tags from a vast and unstructured schema. Experimental results on real-world data from Athens, Greece, demonstrate that a Support Vector Machine (SVM) model using simple geometric and textual features achieves a top-10 recommendation precision of over 92%.

1. Introduction

OpenStreetMap (OSM) is a collaborative project to create a free editable map of the world. While the geometric digitization of entities is relatively straightforward, the subsequent "tagging" process—assigning key-value pairs to describe the entity—is fraught with difficulty. OSM's flexible, non-rigid schema leads to "choice paralysis" for users and data inconsistency across the platform. This research proposes OSMRec, a recommender system designed to suggest the most probable categories for geospatial entities based on their inherent physical and textual properties.

2. Methodology

The recommendation task is framed as a multi-class classification problem. The authors evaluated two distinct algorithmic approaches:

  1. Support Vector Machines (SVM): A model-based classifier that identifies optimal hyperplanes to categorize entities in a high-dimensional feature space.
  2. k-Nearest Neighbors (kNN): A memory-based approach that identifies similar existing entities to suggest categories.

2.1 Feature Engineering

A critical component of the framework is the definition of feature sets:

  • Geometry Type: Categorical flags for entity types (e.g., Point, Polygon).
  • Geometric Metrics: Quantitative data such as area, number of points, and edge length variance.
  • Textual Features: Keywords derived from entity names, processed to remove stop-words and filtered by frequency.
Performance of different feature sets using SVM, highlighting the efficacy of combining simple geometry with textual data.

3. Experimental Setup

The framework was evaluated using a dataset of 111,491 geospatial entities from Athens, Greece, encompassing 1,421 distinct OSM categories. A 5-fold cross-validation approach was employed to ensure the robustness of the results.

4. Results and Discussion

The experimental results indicate that the SVM model significantly outperforms the kNN approach across all precision metrics.

Comparison of SVM and kNN algorithm performance on validation and test sets.

4.1 Feature Synergy

As shown in Figure 1, the combination of "Simple Geometry & Text" provided the best performance. While textual features are highly informative, they are often sparse in OSM data. Conversely, geometric features are always available but less descriptive on their own. The synergy between basic geometric types and available textual keywords allows the model to achieve a Top-1 precision ($P^1$) of 79.04% and a Top-10 precision ($P^{10}$) of 92.07%.

4.2 Algorithm Comparison

The SVM's superior performance (Figure 2) suggests that the relationship between an entity's physical characteristics and its category is complex and better captured by a trained model than by simple similarity metrics used in kNN.

5. Conclusion and Future Work

OSMRec demonstrates the feasibility of automated tag recommendation in crowdsourced geospatial data. The implementation of this system as a JOSM plugin provides immediate practical value to the OSM community. Future research will focus on addressing textual sparsity through contextual features (e.g., spatial proximity) and improving the handling of rare categories in the "long tail" of the OSM schema.

Summary updated

BibTeX
@inproceedings{giannopoulos2015automatic,
  author = {Giorgos Giannopoulos and Nikos Karagiannakis and Dimitrios Skoutas and Spiros Athanasiou},
  title = {{Automatic Recommendations of Categories for Geospatial Entities}},
  booktitle = {LocalRec@RecSys},
  year = {2015},
}
Figure