English. We propose a hierarchical semantic representation of urban areas extracted from a social network to classify the most predominant land use, which is a very common task in urban computing. We encode geo-social data from LocationBased Social Networks with standard feature vectors and a conceptual tree structure that we call Geo-Tree. We use the latter in kernel machines, which can thus perform accurate classification, exploiting hierarchical substructure of concepts as features. Our comparative study on three datasets extracted from Milan, Rome and Naples shows that Tree Kernels applied to Geo-Trees are very effective improving the state of the art.
Italiano. In questo lavoro, proponiamo un nuovo modello semantico per la rappresentazione di aree urbane utilizzando dati da social media. In particolare, modelliamo tale informazione con una struttura ad albero che abbiamo chiamato GeoTree. Questa viene utilizzata, in combinazione con un vettore di feature classico, nelle kernel machine per fare classificazione della destinazione di uso delle aree urbane. Abbiamo valutato il nostro approccio su tre grandi metropoli italiane quali Milano, Roma e Napoli. I risultati mostrano come i Geo-Tree, applicati ai Tree Kernel, riescono a raggiungere risultati di molto superiori ad altri modelli attualmente stato dell’arte.
The growing availability of data from cities
We carried out a study on land use prediction of three Italian cities: Milan, Rome and Naples as follows: (i) we divided each city in squares of 200x200 meters; (ii) then, we classify the most predominant land use class (e.g., High Density Urban Fabric or Open Space and Outdoor), assigned by the city administration. The results show that GeoTKs achieve an impressive improvement over state-of-the-art classification approaches based on BOC., i.e., 21.2%, 13.6% and 54.3% of relative improvement in Macro-F1 over Milan, Rome and
2
Previous work has modeled land use classification
by means of different sources of information. For
example, Yuan et al. (2012) built a framework that,
using human mobility patterns derived from
taxicab trajectories and Point Of Interests (POIs),
classifies the functionality of an area for the city of
Beijing. Assem et al. (2016) proposed a
spatiotemporal approach based on three different
clustering algorithms to model the change of
functionality of a city’s region over time. They extracted
features from Foursquare’s POIs and check-in
activities of Manhattan.
Geospatial city areas are described with the popular shape file format, where each shape is a collection of points geo-localized using their coordinates. The latter are provided with the well-known Coordinate Reference System (CRS) WGS84, adopted for the common latitude/longitude geolocation. We use (i) shape files provided by Urban Atlas1, a website providing data for large urban areas (more than 100; 000 inhabitants) and (ii) POIs from Foursquare2.
Cities are divided in small areas associated with a main land use. In total, there are 17 different land use classes defined from the open dataset Urban Atlas 3. We focused on those related to city centers, discarding those less interesting from a social viewpoint, i.e., associated with rural areas such as forests, agricultural, semi-natural and wetland areas and mineral extraction and dump sites. Thus, we selected the following categories: 1https://www.eea.europa.eu/data-and-maps/data/urbanatlas 2https://foursquare.com/ 3https://www.eea.europa.eu/data-and-maps/data/urbanatlas#tab-additional-information (i) High Density Urban Fabric, (ii) Medium Density Urban Fabric, (iii) Low Density Urban Fabric, (iv) Industrial, commercial, public, military and private units, (v) Open Space & Recreation, (vi) Transportation. We collapsed Medium and Low Density Urban Fabric into one single category, ML-Density Urban Fabric as they only have few samples. Land use distribution is very finegrained, making its classification based on POI information very difficult. A trade-off between classification accuracy and the desired area granularity consists in segmenting the regions in squared cells. As each cell can contain more than one land use label, we consider the predominant label as its primary use.
A POI is usually characterized by a location (i.e., latitude and longitude), textual information (e.g., a description of the activity in that place) and a hierarchical categorization that provides different levels of detail about the activity of the place (e.g., Food, Asian Restaurant, Chinese Restaurant). We used POIs extracted from Foursquare, a geolocation-based social network supported with web search facilities for places and a recommendation system. In particular, we extracted 46,731, 43,389 and 7,219 POIs from Milan, Rome and Naples4, respectively. We focused on the ten macro-categories of such POIs5, each one specialized in maximum four levels of detail. 4
In most machine learning algorithms data examples are transformed in feature vectors, which in turn are used in dot products to carry out both learning and classification. Kernel Machines (KMs) allow for replacing the dot product with kernel functions, which directly compute it on the examples, i.e., they avoid the transformation of examples in vectors. The main advantage of KMs is a much lower computational complexity as it does not directly depend on the feature space size.
The most straightforward way to represent an area by means of Foursquare data is the use its POIs. Every venue is hierarchically categorized (e.g., Professional and Other Places ! Medical Center ! Doctor’s office) and the categories are used to produce an aggregated representation of the area. 4For some reasons Foursquare is less popular in Naples 5https://developer.foursquare.com/categorytree We define a feature vector for a grid cell by counting the macro-level category (e.g., Food) in all the POIs that we found in that cell.
Foursquare has its own hierarchy of categories, which is used to characterize each location and activity (e.g., restaurants or shops) in the database. Thus, each Foursquare POI is associated with a hierarchical path, which semantically describes the type of location/activity (e.g., for Chinese Restaurant, we have the path Food ! Asian Restaurant ! Chinese Restaurant). The path is much more informative than just the target POI name, as it provides feature combinations following the structure and the node proximity information, e.g., Food & Asian Restaurant or Asian Restaurant & Chinese Restaurant are valid features whereas Food & Chinese Restaurant is not.
Geo-Tree: we propose a new tree structure, i.e., Geo-Tree, whose nodes and edges among them are subsets of the Foursquare hierarchy (FH). A GeoTree of a grid cell is constituted by a new root node connecting the subtrees of FH rooted in concepts present in the cell. In other words, we connect all the paths of FH starting from grid concepts. Figure 1 shows an example of the FH paths of a cell and the resulting Geo-Tree.
This way, the nodes of the first level, i.e.,
the root children, correspond to the most general
FH categories, e.g., Arts & Entertainment, Event,
Food, etc., the second level of our tree
corresponds to the second level of the hierarchical tree
of Foursquare, and so on. The terminal nodes are
the finest-grained descriptions in terms of category
about the area, e.g., College Baseball Diamond
or Southwestern French Restaurant. For
example, Fig. 2 illustrates the semantic structure of a
grid cell obtained by combining all the categories’
chains of each venue.
GeoTK: given a Geo-Tree, we can encode all
its substructures in kernel machines using TKs.
In particular, we used the Syntactic Tree Kernels
(STKb) with Bag-Of-Words and the Partial Tree
Kernel (PTK)
BOC kernel: to complement GeoTK, we represent a cell also creating a BOC representation, namely we count the macro-level category (e.g., Food) in all the POIs that we found in any cell grid. This way, we generate feature vectors by counting the number of each activity under each macro-category. In order to take into consideration the popularity of the area, we included (i) the total sum of unique users that did at least one check-in in the cell, and (ii) the total sum of check-in done in the cell. Note that, given an area, the number of unique users provides an idea on how many people visited it, while the number of check-in can be used to represent its popularity.
Kernel combination: finally, given two geographical areas, xa and xb, we define a kernel combining Geo-Tree and BOC as: K(xa; xb) = T K(ta; tb) + KV (va; vb), where T K is any 6It is possible to add the frequency in the kernel computation but for our study we preferred to have a completely different representation from previous typical frequency-based approaches. structural kernel function applied to tree representations, ta and tb of the geographical areas and KV is a kernel applied to the feature vectors, va and vb, extracted from xa and xb using any data source available (e.g., text, social media, mobile phone and census data).
We performed our experiments on the data from
Milan, Rome and Naples. We used a grid of
200x200meters as it is indicated as the best size
from other similar previous work on land use
classification
To train our models, we applied
SVM-LightTK7, which enables the use of structural kernels
We trained multi-class classifiers using
common learning algorithm such XGboost
8http://svmlight.joachims.org/ 9https://github.com/aseveryn/SVMTK-MulticlassClassifier nomial and radial basis function kernels, named SVM-fLin, Poly, Rbfg, respectively, and our structural semantic models, indicated with STKb and PTK. We also combined kernels with a simple summation, e.g., PTK+Lin indicates an SVM using such kernel combination.
Table 1 shows the average of F1, Precision and Recall over the different categories. The model baseline is obtained by always classifying an example with the label High Density Urban Fabric, which is the most frequent. Due to space constraint, we only reported six models, namely: the baseline, XGBoost and the top four kernel models.
We note that: (i) GeoTK always outperforms XGBoost and the baseline, demonstrating the superiority of our novel approach. This is an interesting finding as XGboost is the current state of the art for land use classification. (ii) STKb combined with feature vector always produces the best results, improving the F1-score over XGBoost up to 6.3, 3.8 and 11.9 absolute points for Milan, Rome and Naples, respectively. (iii) Kernel combinations always provide the best results.
In this paper, we have introduced Geo-Trees, a novel semantic representation based on a hierarchical classification of POIs, to better exploit geosocial data to the classification of the primary land use of an urban area. This is an important task as it gives the urban planners and policy makers the possibility to better administrate and renew a city in terms of infrastructures, resources and services. More in detail, we have built our classifiers with combinations of a kernel over BOC and TKs applied to Geo-Trees, thus exploiting hierarchical substructure of concepts as features. Our comparative study on three large Italian cities, Milan, Rome and Naples shows that our models can relatively improve the state of the art up to 11.9 absolute points in F1-score.
This work has been partially supported by the EC
project CogNet, 671625