The Web of Data puts a vast and ever-increasing amount of information at the disposal of its users. In the era of big data, interpreting and exploiting these information is both a highly active research area and a key issue for users in industry trying to gain a competitive edge. One current problem in industry with many potential application areas is nding a common theme for varying features by generating higher level summaries. We introduce the notion of motives to describe these common themes. Motives can be identi ed for all sorts of entities such as geo-spatial regions (e.g., \cultural regions") or holidays (e.g., \winter holidays", \activity holidays"). These motives are closer to common language and human conversations than ordinary keywords. Since users prefer formulating their information needs using everyday language, which expresses their understanding of the world, the potential for a strong industrial impact for search applications can be derived. However, capturing the users' often vaguely formulated intentions and matching them to appropriate retrieval operations on the available knowledge bases is a challenging issue. Yet, it is an important step on the way of providing the best possible search experience to users. This paper presents our work in progress on computing motives for geospatial regions. Following a long term agenda, we are evaluating the requirements for identifying such motives in large data sets. At this point, we can show that out-of-the-box machine learning methods can be used on Linked Data to train a model for computation of geo-spatial motives with good accuracy.
One approach is to annotate all available pieces of information with a
probability representing the con dence in this information, e.g., like it is done in
the Google Knowledge Vault [
A well-known methodology of structuring data is to model categories and
assign them to data items of the knowledge base. For example, for structuring
geo-spatial entities, a good category might be \regions where educational
institutions are located". Motives express the existence of a high relevance of the
given characteristics (e.g., \educational institutions"). Hence, a category in the
sense of a motive might be \regions well known for there educational
institutions", i.e., it is a common interpretation of a highly relevant set of entities.
In previous work, this was very successfully applied within search-driven
applications [
One of the crystallization points for the Web of Data is DBpedia [
Therefore, it can be assumed that computing a motive from DBpedia is nontrivial and at least the following challenges need to be addressed: 1. local communities are active with a di erent intensity, e.g., there are 1.193.557 Wikipedia articles written in Polish and just 35.154 in African3 although a similar number of people are speaking these languages, 2. the history of entity types di er with regard to culture and administrative regions of the world, 3. (administrative) regions have varying de nitions depending on the country they are in.
However, the Web of Data gives us the opportunity to use further, interlinked
data sets to increase the number of available features for the considered entities.
This contribution uses Natural Earth data [
Additional problems appear when working with globally crowdsourced Linked Data: 1 manual annotations are needed for the training 2 http://www.wikipedia.org 3 Fetched from www.wikipedia.org on 2015-04-20. 1. the interpretation of data might di er from region to region, e.g., the number of educational institutions in the United Kingdom is way higher than the one in France (following DBpedia), although the importance of higher education can be considered as equal in both countries and the number of inhabitants is close to equal, 2. the data is incomplete, i.e., it cannot be assumed that all entities of a type are captured, 3. there is no certainty property available.
Calculating motives or determining their presence is hence a challenging task, as humans, naturally, will judge calculated motives by their experience, which is in uenced by many factors such as cultural background, place of residence or the education level. For this paper, the presence of motives was judged by an expert committee. In an iterative process, we used machine learning techniques to train the properties (i.e., features describing the geo-spatial motives).
The paper is organized as follows: The next section presents related work. Section 3 describes data sets used in this contribution. The idea and requirements for motive computation are shown in Section 4, while a case study is performed in Section 5. Finally, the paper is concluded in Section 6, where also the future work is described. 2
The vision of a machine-processable Web of Data originating from Tim
BernersLee has been described in its aspects and principles in [
While some authors point out a lack of research at the intersection between
Linked Data and machine learning [
Furthermore, nding abstractions and grouping entities in DBpedia has been
a subject of interest close to our contribution. In [
DBpedia is a crowd-sourced community e ort to extract structured
information from Wikipedia5 and make this information available on the
Web. [
Natural Earth is a public domain map data set (. . . ). Featuring tightly
integrated vector and raster data, with Natural Earth you can make a
variety of visually pleasing, well-crafted maps (. . . ). [
Hence, the main purpose of Natural Earth is to provide data required for
computing maps and their visualisation. However, the data is linked to GeoNames
[
The GeoNames geographical database covers all countries and contains
over eight million placenames that are available for download free of
charge. [
In this section, we will describe the properties of geo-spatial motives and collect the requirements for their computation.
A motive can be de ned as the reason for a search as well as particular
conditions like how much a product should cost. [
Accordingly, it can be assumed that a motive for a populated place exists if potential users will associate it with the related concept. We express the relation by the following de nition: De nition 1 (Geo-Spatial Motives). A geo-spatial motive P is named as property urn:unister:has-geospatial-motive and O is an entity expressing the motive. Hence, the triple S P O expresses the relation, where S has a property of the type http://www.georss.org /georss/point.
That is, the following fact might exist with regard the example from above: dbpedia:Innsbruck urn:unister:has-geospatial-motive urn:unister:-geospatial-motive:winter-holidays.
As the common characteristics of geo-spatial regions as perceived by users should be represented by a geo-spatial motive, we demand: Requirement 1 (Requirements) A motive m should be bound to a geo-spatial entity e if and only if m is a commonly associated characteristic of the considered region e.
Hence, with regard to our available data sets (c.f., Section 3), geo-spatial DBpedia entities might be annotated with their type as motive if they are well-known within the considered country. 5
In a case study, we will show how di erent data sets containing di erent features in uence the quality of a trained model for the computation of motives. For the evaluation, we choose the education motive as it is present in most of the countries and can be interpreted for countries without being part of the local community. That is, populated places should be annotated with this motive if and only if they are well-known for their educational institutions with regard to the country they are located in. This limits the selection of relevant DBpedia entities to entities of the type http://dbpedia.org/ontology/EducationalInstitution, hereafter called educational entities. The extracted entities are shown in Figure 3 to give a visual impression. Moreover, populated places are considered if and only if there is at least one educational entity nearby10. 5.1
For our experiments, we used DBpedia, Natural Earth's populated places and countries, as well as the GeoNames data set. In addition, we annotated several new features that should be useful for detecting the education relevance of populated places. First, using DBpedia, we calculated the number of educational entities within a radius x = 10, 25 and 50 km for every populated place (these features will be further referred as entities nearby xkm). Adding information from the Natural Earth data set, which contains states and countries in which the populated places lie, it is possible to derive more sophisticated statistics: { the maximum of the number of educational entities nearby x km over populated places in a country (referred as max of places within xkm) { the maximum of the number of educational entities over states in a country (referred as max of states) 10 The maximal radius of the considered bu er around a populated place is de ned by 50 km derived following an educated guess. { the number of educational entities in the country (referred as e in country)
Figure 4 presents histograms that illustrate the number of populated places with a certain number of educational entities within the radius of 10, 25 and 50 km. As one can see, almost all places have up to 10 20 entities within 10 km, whereas the distribution of places with entities within 50 km is more at. Thus, many places have relatively few educational entities nearby, which could be an evidence for their low education relevance. However, as the Wikipedia community in some countries might be not very strong, it is also possible that not all entities are represented within the data set (we leave this for later evaluation). The joined data set of populated places was labeled by several quali ed experts. Given the name of a populated place, its state and the country, they classi ed the place as \highly relevant" (ranking r = 2), \intermediately relevant" (r = 1) or \irrelevant" (r = 0) for the education in the country of this place. Up to now, we employed three experts in the age range of 30 38. Two of them have doctoral degrees in philosophy and mathematics, respectively. In addition, they have working experience in academia. The third person is currently a member of the international Masters programme in archaeology and has performed studies at several international universities. Therefore, we assume that their quali cation helped them to label data as objectively as possible.
The experts were provided with a data set of 2000 places from 171 countries. They were free to rate only places which are known to them. As a result, there are 881 labeled data samples from 121 countries available for our experiments. As data classi cation by experts is time consuming, in the future we might infer place labels using various ranking lists of educational institutions. 5.3
The main questions under investigation are twofold: 1. Can one use machine learning techniques on features of linked data for computing motives? 2. Is it bene cial to aggregate data from di erent sources? We investigated these questions by building classi ers for motive computation and comparing their accuracy while iteratively increasing the number of data sets and features used.
In the rst step, the expert ratings were merged. The resulting rating of a
place is just the average of all available ratings giving a continuous value [0; 2].
As the number of labeled populated places is rather small, the ratings were
converted into binary values. Thus, a place can have a rating of either \relevant"
or \irrelevant" for the education in its country, corresponding to the initial ranges
[0; 1) and [1; 2], respectively. In this way, the classi cation problem is simpli ed
to account for the limited amount of training data. Data preprocessing and
classi cation were done in Weka [
As baseline for our experiments, we used existing and annotated DBpedia
features of populated places. Of course, this data set is at least required to make
a statement about the educational strength of a place. The second data set
was formed by merging the extended DBpedia data set with data from Natural
Earth's populated places and countries as well as features that we calculated
using this data (see Section 5.1). Finally, the third data set extends the second
one by GeoNames attributes of populated places. These data sets will be further
referred as D1, D2 and D3, respectively. Thus, all three data sets contain 881
populated places which are described by di erent features. Table 1 presents
features that were included in every data set, see Table 3 for the meaning of used
Natural Earth and GeoNames features. These lists are not complete and contain
only features that were regarded as relevant for classi cation by the
correlationbased feature selection algorithm [
As a classi er, we picked Breiman's random forest [
Table 2 presents evaluation metrics of classi ers build on di erent data sets (Table 4 gives the detailed explanation of the analyzed metrics). All metrics are 11 c.f., http://www.naturalearthdata.com/downloads/10m-cultural-vectors/10mpopulated-places/ averaged over 200 runs. For every run, the original corresponding data set with 881 samples was randomized and divided into a training and a testing set containing 66% and 34% of the original data, respectively. Despite the fact that the classi cation accuracy is not perfect, in our opinion, the results are promising. One can clearly see the classi er built on D3, the data set containing DBpedia, Natural Earth and GeoNames data, achieving the best results. Moreover, enriching DBpedia data only with the features from Natural Earth improves the classi cation performance drastically. Though, the di erence between D2 and D3 is not large, all metrics achieved on D3 are signi cantly better that those achieved on D2 according to the Wilcoxon signed-rank test at the p-level= 0:05. Adding just a single feature from GeoNames, gn pop, it is possible to achieve the higher values of the accuracy and the F-measure, which illustrates the general classi er performance and its performance w.r.t. the positive class only, respectively. This means that using features of D3, one can classify both educationally relevant and irrelevant populated places more precisely. Thus, one obviously pro ts from merging various data sets while trying to automatically detect motives. Many data sets, if merged properly, provide a large pool of features that can be useful for detecting di erent motives.
Detailed evaluation results of the classi er trained on D3 are presented in Table 6 in the appendix. The table contains metrics of 5 runs for every class separately and their weighted average. Note that training data is highly unbalanced having 30% educationally relevant and 70% irrelevant populated places. This fact explains the di erence in values of the evaluation metrics for two classes.
Note that the feature selection process, which was run on D2, selected our annotated features as relevant for classi cation. Moreover, it seems that information over all distances, 10, 25 and 50 km, is relevant, which might account for di erences between regions with the diverse population density. Therefore, we assume that including more features concerning local geo- and demographic statistics will help to improve the classi cation accuracy. 6
Making knowledge bases searchable is a key challenge considering the data economics. Hence, both academics and industry have to increase their e ort to make knowledge bases accessible for actual (human) users. Geo-spatial motives are one approach to enable end-users to connect their common sense of the world knowledge with the actual representation within the knowledge base.
In this paper, we presented and evaluated geo-spatial motives, an approach for a more general and exible representation of the data available in knowledge bases. With these geo-spatial motives, we will push the available data towards better understandability and aim for better search-driven applications.
We have shown that machine learning techniques on Linked Data can be used for computing geo-spatial motives. Moreover, it seems crucial to include the knowledge of di erent data sets for good quality of the calculated motives. Despite the di erent challenges triggered by the properties of the available data sets, it was possible to achieve a promising motive detection quality within the considered case study. Therefore, we can assume that industrial search-driven applications can take advantage of knowledge bases in the Web of Data to provide a more human-friendly interaction.
However, this paper is just one step on our research agenda working towards knowledge bases that can be used and be of bene t for both experts and nonexperts. In the future, we will evaluate the capabilities of di erent knowledge bases and generalize our approach with the aim of preventing a training for all expected motives. Instead, we will establish mechanisms that work by analogy. Of great impact might be the integration of di erent levels of geo-spatial regions like admin-1 entities (states, provinces) of Natural Earth12 to increase the quality and to extend the considered motives from populated places to natural regions like the Alps (e.g., winter sports) or the Silicon Valley (e.g., IT companies). Finally, it has to be evaluated how the di erent personal backgrounds of users (e.g., education level, cultural background, age) is in uencing their perception of motives.
Acknowledgments We thank Luise Erfurth, Stephan Schwinger, Bernd Eickmann and Kristin Mittag for their valuable support. This work has been supported by grants from the European Union's 7th Framework Programme provided for the project GeoKnow (GA no. 318159). 12 http://www.naturalearthdata.com/downloads/10m-cultural-vectors/ 10m-admin-1-states-provinces/ Shortcut Meaning worldcity indicates whether a city is important for the global economics pop max the population of the city metropolitan area rank max a population rank calculated based on pop max max aremi a metropolitan area of the city in squared miles
gn pop the city population according to the GeoNames
Class TP Rate FP Rate Precision F-Measure AUC Accuracy relevant 0.514 0.097 0.627 0.565 0.842 0.809 irrelevant 0.903 0.486 0.854 0.878 0.842 0.809 weight. avg. 0.809 0.392 0.799 0.802 0.842 0.809 relevant 0.513 0.117 0.6 0.553 0.815 0.789 irrelevant 0.883 0.487 0.842 0.862 0.815 0.789 weight. avg. 0.789 0.392 0.780 0.783 0.815 0.789 relevant 0.494 0.086 0.672 0.569 0.826 0.803 irrelevant 0.914 0.506 0.834 0.872 0.826 0.803 weight. avg. 0.802 0.395 0.791 0.792 0.826 0.803 relevant 0.467 0.107 0.593 0.522 0.804 0.786 irrelevant 0.893 0.533 0.833 0.862 0.804 0.786 weight. avg. 0.785 0.426 0.773 0.776 0.804 0.786 relevant 0.488 0.082 0.684 0.569 0.849 0.803 irrelevant 0.918 0.513 0.831 0.872 0.849 0.803 weight. avg. 0.802 0.397 0.791 0.791 0.849 0.803 0.797 0.400 0.787 0.789 0.827 0.798