In this paper we propose the use of an ontology matching algorithm to guarantee the interoperability of the different agents that integrate an smart city. In this sort of environment the different parties need to cooperate and to integrate their information in order to provide enhanced services to the users of the smart city. As the information of these parties may be described by means of different and heterogeneous ontologies, we find the solution in the use of ontology matching techniques. The algorithm presented was designed to be able to exploit the knowledge of previous matched agents to enhance the results and provide the most accurate results possible.
In the last years there has been a remarkable increase in the amount of projects and
initiatives related to Internet of Things [
The idea behind integrating these devices in a city is to turn it into a smart one, so
citizen’s lives can be improved with new types of services and comfort. These services
can be related to almost every aspect of city life and infrastructure, water and energy
supply, transportation, healthcare, education, etc. [
To turn a city into a smart one, the first task to address is to develop a rich
environment of networks that support digital applications [
Hence, in a smart city, smart devices, Sensor Networks (SNs) [
It is highly likely that the deployment of a smart city is not done all at once but in a
series of steps, so it is equally likely that different parts of the smart city are developed
by different parties, resulting in the coexistence of different public and private
deployments, each one of which possibly using different smart devices and also their
own hardware and software architectures. It is necessary to guarantee the success of a
smart city to put a special interest in allowing that these different deployments will be
able to interact and seamlessly communicate with each other and, that the information
gathered by the different devices will be properly integrated and shared among the
different systems [
This problem is not new to the research community and several alternatives have
been already proposed [
Our proposal includes the use of ontology matching techniques [
The remaining of the paper is organized as follows. In section 2, we delve into the use of ontology matching in smart cities and provide the foundations that supported the development of our system. In section 3, a description of our solution is provided and discussed. Finally in section 4, the main conclusions and future lines of work are summarized. 2
A smart city may be seen as a distributed system where several agents on behalf of
their users collect data from the environment by using different sensors. The concept of
user here should be globally understood, as the user of an agent may be a citizen, a
smart device, an application, another agent, etc. The use of ontologies in smart cities is
not new as there is for instance the SCRIBE ontology [
However, there are several reasons why ontologies by themselves are not enough to
guarantee the interoperability of the different agents. For instance, the agents may use
different ontologies to represent the information gathered from the sensors, the
software applications in the smart city may be developed by different providers that
represent their internal knowledge using different ontologies, there may be agents or
applications included in the smart city in a later stage or even itinerant agents that only
need a concrete service at a certain time. In order to actually reduce the heterogeneity
in the definitions and allow a seamless communication of the parties, we relied on
ontology matching techniques [
These techniques allow the identification of alignments for pairs of ontologies
where an alignment identifies the set of correspondences holding between the entities
belonging to the ontologies [
Different classifications have been made for the matching techniques although for
the scope of this paper, we followed the one that Euzenat and Shvaiko propose in [
In the following section, while describing our solution to the ontology matching problem in smart cities, we briefly describe the different techniques that we have used linking them to this classification. 3
In this section we briefly describe our algorithm for ontology matching in smart
cities and how we have enhanced its results by following an alignment reuse [
It takes as input two OWL [
These initial correspondences are obtained by using some language-based and terminological techniques. The language-based techniques consider names as words in a natural language and exploit their morphological features. Some of the methods used, as part of the pre-processing of the strings, are tokenisation, that consists of splitting words into shorter sequences by means of a separator (blanks, punctuation marks, camel-case changes, etc) and stopword elimination, that consists of removing words such as articles, prepositions, etc.
On the other hand, the terminological techniques consider their inputs just as strings
and apply string-distance measures to asses the similarity between two entities. In our
case we have used Jaro-Winkler distance [
Once the similarity between the terms in the ontologies has been determined, only those pairs with the highest value are selected to become the initial binding points.
These initial correspondences sequentially undergo several procedures that take advantage of some structural features of the ontologies and that allow the discovery of new binding points. These binding points can identify both pairs of classes or properties. Each one of the newly discovered binding points is assigned a tag that identifies the procedure and branch within it that led to its discovery. If a binding point is reached by several procedures, all the tags are recorded. 1. Properties Inverse Procedure: this procedure retrieves new correspondences between properties by exploiting the existence of inverse properties defined with the construct owl:inverseOf. 2. Properties Domain Range Procedure: this procedure obtains new correspondences between classes by comparing the domains and ranges of the initial properties. Not only the first-level domain and range classes are evaluated but the procedure continues until reaching the higher levels of the hierarchy. 3. Classes Properties Procedure: this procedure allows the retrieval of both new correspondences between classes and properties. This procedure recursively identifies the similar properties existing among the class correspondences, and then assesses the existence of other classes belonging to the domain or range of this properties that could be a new correspondence. 4. Classes Family Procedure: this procedure retrieves new correspondences between classes. It exploits the familiar relations of the classes. For each pair of them, its superclasses, subclasses, and sibling classes are evaluated to determine the existence of new possible matches.
These procedures are iteratively applied until no new correspondences are discovered. Once these procedures have finished all the correspondences that have been discovered are filtered to produce the final output of the algorithm. To do so, the tagging is very important as it allows the identification of the different procedures and sub-procedures. It is based on the idea that the different procedures exploit different structural features of the ontologies and hence the likelihood that the obtained results are good is not the same for all of them.
To evaluate the performance of the algorithm we intended to use ontologies from
the smart city domain. However, the amount of ontologies in this area proved not to be
enough to allow an accurate evaluation. Hence we have used the testbed provided by
the Ontology Alignment Evaluation Initiative 2013 [
= # # _
_ ─ Recall: measures the ratio of correct correspondences over the number of expected ones. It reflects, the degree of completeness of an algorithm. # # _
_ ─ F-measure: is a measure introduced to compare the systems with just one value since it is highly likely that the system with a higher recall may have a lower precision and vice versa.
= !"∗ ∗ # ! ∗
These measures were used to evaluate the performance of our algorithm. Among the range of tests at the OAEI-13 we have tested our algorithm with several of them, although for the scope of this paper we will be focussing on the conference track which aims at finding alignments within a collection of ontologies from the domain of conference organization. The results obtained by our algorithm for each pair of input ontologies are compared with a reference alignment also obtained from the OAEI-13 website. In table 1 we include the average results obtained for this task.
In the smart cities domain, there is a series of ontologies that describe the resources and services that are available for the agents. If an agent needs a certain resource or service, it will need to match its ontology to the appropriate one in the smart city. Depending on where the service or resource is deployed, the agent may need to match its ontology to a part of the ontology that describes the smart city itself, usually when the agent needs access to a resource, or to another agent’s ontology, usually when the agent needs a service that is offered by the other one. This situation is depicted in figure 2. In any case, this process will output an alignment between both ontologies. If several agents need to access the same resource or service, the process will be repeated several times.
Our intuition is that if a new agent arrives in the smart city and is willing to use a service or resource, the alignments previously obtained from other agents may help in tuning the alignment process for this new agent and therefore they may be used to enhance the results produced by the algorithm.
This led us to delving into alignment reuse techniques [
To asses the viability of integrating alignment reuse as part of the ontology matching proposal for smart cities, we have used the ontologies of the conference track. The procedure followed to do so is to feed the algorithm with some intermediate alignments that are then used to identify binding points between the ontologies to match.
Consider the following example, let us suppose that there are three different ontologies, , and , and that we need to match ontology to ontology . If we also have available the alignments between and ( _ _ ) and, and ( _ _ ), then it is possible to identify a path that, using these intermediate alignments, may link entities in to entities in . We refer to this as a ring between and through , and it is graphically represented in figure 3.
Tables from 2 to 6 show the results of testing this approach with the ontologies of the conference track. From the ontologies available at this track in the following examples we have used the following ones: cmt, conference, confOf, edas, ekaw and sigkdd.
Table 2 shows the results obtained by directly matching the conference ontology to the confOf ontology. These values are included to provide a baseline to compare the results obtained when using rings.
When using edas for the ring, 10 different paths from conference to confOf were detected which allowed the identification of 5 new correspondences. Using ekaw, just 8 different paths were identified which added 2 new correspondences that were not detected when directly running the matching process. We considered then a combined approach using both edas and ekaw at a time, seeking to obtain results with the precision enhancement provided by ekaw and the recall enhancement provided by edas. The results obtained are shown in table 5.
Using a multiple ring the amount of identified paths rises to 13. The results obtained with this approach show that precision is not as high as when just using the single ring with ekaw as there is an extra incorrect correspondence that is added to the final output. However, for recall and F-measure, the values obtained are remarkably higher that those obtained with single rings and when compared to the baseline results we observe an improvement of 10.95% for precision, 62.26% for recall and 35.48% for f-measure.
In spite of being a positive outcome, the results vary when using the alignments with other ontologies as rings. Table 6 shows the results obtained when considering the alignments with cmt and sigkdd for the rings. The results in this table also show an improvement compared to the baseline in table 2, although they are not as remarkable as those in table 5.
Other tests run using more alignments showed no improvement compared with using just two as in the examples presented previously. However the testbed that we used is not large enough to entirely dismiss that possibility. An issue that we have identified is that even when using the reference alignments provided by the OAEI, which are the golden standard used to compare the results of any algorithm, there were some paths that we identified, that led to selecting as binding points pairs of entities that then were not considered as a valid correspondence in the reference alignment. 4
In this paper we have introduced the smart cities domain and underline some communication and interaction problems that are highly likely to show up in this kind of development. We have also described the foundations of the ontology matching based approach that we propose to tackle such problems in the smart cities, and the ontology matching algorithm that we have defined to address this problem. We have described the measures of precision, recall and f-measure used to evaluate this type of algorithm, and the results obtained when doing so.
We have then looked at some alternatives to refine the alignments obtained by our algorithm and therefore improve such results. Among the different techniques we have focused on alignment reuse as it was particularly indicated for scenarios like ours. We have used previously existing alignments, rings, to identify paths of between the ontologies to match and therefore enhance the results obtained. We have tested this approach using the ontologies available from the conference track of the Ontology Alignment Evaluation Initiative 2013. And we have verified the viability and validity of the approach.
In spite of these good results that account for the viability of our approach, there are some issues that need to be addressed in order to obtain the best results possible and hence to improve the usability of the smart cities. There is, for instance, the need to test our proposal using ontologies taken from the real domain where it will be deployed, the smart cities. Additionally, as we introduced in section 3, there seems to be a direct relation between the rings chosen and the goodness of the results obtained, so we aim at focusing on determining the features that make a ring better than other. It is also necessary to explore techniques that will allow the alignment reuse in the real environment, so we are turning to some techniques such as alignment storing and sharing and alignment annotation.