RDF data has been extensively deployed for the interlinking of health-related data in a structured way. The de nition of link statements between distinct resources plays a key role to interconnect several life science repositories. However, RDF assertions are subject to change, which can a ect existing links. In this article, we conduct extensive experiments to understand the evolution of links in the Linked Open Data (LOD). The objective is to empirically associate changes in the semantic de nition of data resources with modi cations observed in prede ned links. We consider two versions of the Agrovoc RDF repository to calculate di erent types of change operations and associate them to link change actions. Obtained results indicate the existence of the cases investigated in this study. We demonstrate that RDF changes impact the evolution of established links.
Linked data technologies have become increasingly relevant for semantic interoperability and knowledge discovery in the life sciences. Their data sources are more and more interconnected via links between Resource Description Framework (RDF) data descriptions. The interconnection of RDF statements via explicit links plays a central role to assure data linkage. Datasets focusing on life sciences, such as Bio2RDF, Mesh and Agrovoc are part of the LOD cloud. These datasets are subject to a constant evolution via the publication of new releases.
In order to couple with domain updates, RDF statements de ning real-world
resources are subject to change. RDF triples can be updated, added or removed
to keep the repositories up-to-date. Although the implementation of change
operations in RDF datasets is essential to assure structured data evolution, these
operations can negatively a ect established links. This hampers data linkage
consistency over time. Indeed, the connection between nodes can be broken by
a number of reasons, which changes the state of the link to broken or invalid [
Literature has super cially studied the maintenance of linked datasets [
In this article, our thorough and original experiments analyze the way change operations in RDF repositories correlate to changes observed in links. We systematically investigate the key factors of link evolution according to RDF change operations. We assume that the results of our analyses must allow understanding and predicting the consequences of changes performed in RDF datasets over links. We aim to understand the necessary elements regarding the changes detected in the bases to adapt links between instances of two datasets, i.e., external links from one dataset to another. This study might support the de nition of automatic mechanisms for link evolution in LOD.
Our experiments consider change operations in RDF repositories and the modi cations observed in the level of links. This investigation performs several analyses to measure the extent to which the change operations are associated to the observable changes in links. For instance, how the removal of internal triples a ect existing links to external datasets. We conduct experiments with versions of the Agrovoc4 linked dataset.
Our ndings demonstrate that changes in links (like addition, removal and modi cation) are caused as e ects of modi cations in predicate or object of internal triples. We found that in most cases that an internal triple is added, a link is created. However, when an internal triple is removed, only few cases of links associated to the triple is removed, leading to the problem of broken links.
The remaining of this article is organized as follows: Section 2 presents the related work; Section 3 reports on the description of the experimental study; Section 4 presents the obtained results, which are further discussed in Section 5. Finally, Section 6 draws conclusion remarks. 2
The evolution of ontologies and mappings between them have been studied over
the last years. Dos Reis et al. [
There has been a number of studies on speci c aspects of LOD evolution. Literature has addressed techniques to track changes in RDF repositories for the
automatic identi cation of deltas between versions [
Concerning the RDF link aspect, few studies approached the link integrity
problem aiming to monitor and preserve data quality [
Liu and Li [
Studies have addressed the broken link problem in the context of LOD.
Popitsch and Haslhofer proposed a event detection framework (DSNotify ) [
The framework MeLinDa [
In contrast, the Diachron platform [
By using links with predicates as \same as" and \see also", Vesse et al. [
Silk [
Very few studies have attempted to empirically understand the dynamics
of Linked Data. Kafer et al. [
Although these studies present valuable attempts to somehow reach link maintenance, they fail in supporting the link evolution and remain unable to fully tackle the impact of the RDF changes in link evolution. In this research, we carry out empirical studies contributing to pave the way to the de nition of automatic techniques of link evolution in LOD. Our long-term goal is to track the behaviour of links in the datasets in order to keep the links up-to-date as automatically as possible, minimizing the occurrence of broken links. 3
This study aims to comprehend how changes related to RDF facts impact the
evolution of links in LOD. We rely on the main building blocks of the Semantic
Web paradigm, referring to the Universal Resource Identi er (URI), RDF and
ontologies [
RDF dataset. A dataset in the context of Linked Data is an conglomeration of a nite number of RDF triples in a domain. Formally, R = ft1; t2; t3; :::; tng.
Triple. A RDF triple refers to a data entity composed of subject, predicate and object de ned in the form of t = (s; p; o) where: { Subject: (s) is either an URI reference or a blank node. { Predicate: (p) is a URI reference that de nes a relation between a subject and a predicate. { Object: (o) is either an URI reference, a literal, or a blank node.
A literal is a string combined with either a language identi er (plain literal) or a data-type (typed literal). Blank nodes are those nodes representing the resources for which a URI or literal are not given.
Ontology. An ontology O describes a domain in terms of concepts (general
understanding of something), attributes and relationships [
Link. Besides the use of triples inside a dataset, the linkage among several datasets interconnect the datasets. To this end, it is necessary that a predicate establishes a relation between a subject in the rst dataset (source) and an object in the second (target). Formally, we de ne a link as l =< ra; p; rb > connecting a pair of resources ra and rb, in which ra 2 RS and rb 2 RT , such that RS di ers from RT . For the de nition of p, we consider well-established properties to express the predicates of links including: owl : SameAs, rdf s : seeAlso, owl : Dif f erentF rom and SKOS mapping properties vocabulary5. This study relies on the notion of the following types of links, consisting of predicates connecting the Agrovoc datasets with other datasets: { l1 =< ra; owl : sameAs; rb >
{ l2 =< ra; rdf s : seeAlso; rb > { l3 =< ra; owl : dif f erentF rom; rb > { l4 =< ra; skos : exactM atch; rb > { l5 =< ra; skos : closeM atch; rb > { l6 =< ra; skos : broaderM atch; rb > { l7 =< ra; skos : narrowM atch; rb >
From now on, we use the notation l(ra ! rb) to denote a link. We de ne a set of links between RS and RT as LST = fl0; l1; l2; :::; lng. A complete dataset refers to the union of internal triples and links, such as D = R [ LST .
Link change actions. Our study was structured to understand the link evolution in the LOD. To this end, we assume that there might exist di erent types of changes in links. Given two versions of the same RDF repository, lk a link, we de ne the link change actions as follows: { Unchanged link (lk 2 jLST j ^ lk 2 jL+S1T j+1) { Added link (lk 2= LST ^jlk 2 LST j)+1) { Removed link (lk 2 LSTj ^ lk 2= LSTj+1, pj 6= pj+1 or rbj 6= rbj+1 ) { Modi ed link (lk 2 LST ^ lk 2 LST
RDF change operations. Considering two di erent versions of the same
dataset, it is necessary to compute di erences between RDF graphs. At this
stage, we introduce the notion of time j 2 N . For a RDF dataset, we denote
oRpSejratthioenisniatirael rveeprrseiosenntoefdthbeydaantaosepteraantdorRS,j+s1u,cthhetheavtolve(dRvSejr;sRioSnj.+1C)h(aanlgsoe
known as a Di operation) produces low-level changes as triple additions and
removals in the RDF graph [
Complex changes refer to a set of well delimited simple changes; CCH = fSCH1; SCH2; :::; SCHkg. A combination of simple change operations is necessary to obtain a complex change. For example, the replacement of a resource is the combination of the removal of a triple with the addition of a new one.
The evolution of RDF datasets in terms of changes a ecting their triples
may invalidate previously determined links. RDF statements (triples) remain the
smallest manageable piece of knowledge. Modi cations including the addition,
removal, or update of a resource might a ect triple statements, which can lead to
a link integrity issue. In order to maintain the consistency of RDF datasets, its
links should remain in a integrity state, even with underlying changes in data [
Consider a link lj at time j and a link lj+1 based on distinct releases of the associated datasets. Modi cations occurring in related resources of the link (ra) from a release of the dataset in a speci c time j to a time j +1 can invalidate such link lj(ra ! rb). For example, ra can exist at time j (ra 2 RSj ), but be removed at time j + 1 (ra 2= RSj+1 ). In this sense, the link is considered structurally broken and lj should be updated to a new state lj+1(ra ! rc), such that, in this case as an example, ra 6= rc.
In this study, for each change in a ecting triples of a RDF dataset associated to links, we investigate the way changes occured in links. To this end, we selected two distinct versions of the same RDF dataset. Our study explored the Agrovoc6, a well-known dataset in life sciences related to agriculture, food and environment. Agrovoc's maintainers regularly publish their open data on the web. Our criteria for such selection was considering a well-known dataset with characteristics of evolving over time. From one release version Rj to another Rj+1, we calculated ontology change operations (RSj ; RSj+1 ). We used this information to correlate with changes in the links LST .
Table 1 presents the releases of the datasets used in our investigation. Dj corresponds to the release before evolution, and Dj+1 stands for a new release of the same dataset. { The total number of triples is 4.254.655 in the initial dataset release Dj and 4.540.205 in the newer release Dj+1; { In Dj , there are 64.977 links (also named here external triples) - divided into \closeMatch", \exactMatch", \narrowMatch", and \broadMatch" predicates - representing 1.52% of the whole dataset; { The (DSj ; DSj+1 ) (changes computed between the two dataset releases) consists of 535.467 changes, such that: 202 removals (0.03%), 6.990 additions (1.3%) and 528.275 modi cations (98.65%);
We de ned three types of analyses to investigate the evolution of the links. The rst focuses on the relation of removed triples with links; the second emphasizes the behaviour of added triples; the third aims to map the relation between modi ed triples. For all these analyses, we assume that only one side of the dataset evolves per time. In this sense, we can assure that the observed e ects in the links occur due to the observable modi cations from one version to other of the RDF repository.
Analyses of removed triples over links. Figure 1 (a and b) presents the cases related to the removal of triples from one version to another. { Case R:1 (Figure 1 (a)). In this case, there is a link lj (ra ! rb) and the triple associated to the resource ra is removed. In this sense, the resource
ra 2 tk such that tk 2 RSj and tk 2= RSj+1 . This analysis searches for a link removal associated to the triple tk. We consider that exists a coherence between the removal of the triple and the link. { Case R:2 (Figure 1 (b)). In this case, a triple tk is removed from the RDF dataset RSj to RSj+1 , but the analysis detects an operation of unchanged link related to the triple tk. We consider this case an inconsistency because a broken link is created due to the dataset evolution.
Analyses of added triples over links. Figure 1 (c and d) presents the two cases related to the addition of triples from one version to another. { Case A:1 (Figure 1 (d)). A new triple tk is added in the RSj+1 ; as a consequence, this analysis aims to identify new links associated to the new triple as a coherence in their evolution. { Case A:2 (Figure 1 (c)). This case presents the scenario in which a new triple is added, but no link is inserted from one version to another of the dataset associated to the new triple. Analyses of modi ed triples over links. Figure 2 presents the four cases related to the modi cation of elements in triples from one version to another. { Case M:1 (Figure 2 (e)). It presents the scenario in which based on a change in the predicate or object of a triple, the analysis recognizes an added link action in the dataset. { Case M:2 (Figure 2 (f)). In this case, a triple tk has its predicate or object changed. This analysis aims to detect a removed link action associated to this change in the internal triple. { Case M:3 (Figure 2 (g)). This case investigates the behaviour of modi ed predicate or object of triples and the modi cation of a predicate or object in the link (triple and link share the same subject). { Case M:4 (Figure 2 (h)). It investigates the occurrences of modi ed predicate or object of triples that did not trigger modi cation in the link associated.
This section presents the results organized by the three types of analyses. Firstly, subsection 4.1 reports on the results of the analyses concerning the way removed triples a ect links; subsection 4.2 reports on e ects of triple addition on links, and subsection 4.3 reports on the e ects of modi ed triples on links. 4.1
Results of the e ects of removed RDF triples
Case R:1. We detected 78 removed internal triples. The total triples present in DSj , but removed in DSj+1 is 202, in which 124 (61.38%) refer to links and 78 (38.61%) are internal triples. We found that 3.85% (i) of the removed triples (a) is also related to the removal of a link. This number found in (a) also represents 1.49% of the total triples removed (j) and less than 0.01% of the total of triples found in the dataset before evolution DSj (k).
Case R.2. The search of (b) returned 75 occurrences of the removal of internal triples, without the removal of associated links, representing 96.15% of internal triples removed (i), 37.13% of both internal and external triples removed (j), less than 0.01% of the total of triples in the dataset before evolution DSj (k). Figure 3 presents an example, in which the triple related to the resource `Lavandula' (represented by the ID `c 4228' in Agrovoc dataset) is removed, and the link connecting `Agrovoc' with an external dataset named `Chinese Agricultural Thesaurus'7 using the predicate `broadMatch' is not removed.
We investigate the scenario in which added triples are related to new links associated to them. Table 3 presents the encountered results of our analysis. The columns entitled Case A:1 and Case A:2 represent the scenarios referred in cases (c) and (d) described in Figure 1, respectively. Case A:1. We found out 6909 internal triples added in DSj+1 that has the same subject of other 494 subjects categorized as external triples (links). Figure 4 presents an example of this case, in which a new triple related to `Lycopersicon' (represented by the ID `c 4474' in Agrovoc dataset) is added, as long as a link to the external dataset named `DBPedia'8 with the predicate `closeMatch' is also created. We found 6990 (e) triple additions in DSj+1 , which 494 (7.07%) (c) are links and 6496 (92.93%) (d) are internal triples. We found that around 106.35% (h) of the internal triples added (a) produced an addition of a link. This number of occurrences that is higher than 100% means that an internal triple can be related to one or more links. This number found in (a) represents 98.84% of all triples added (i), and 0.15% of the total of triples in DSj+1 (j) .
Case A:2. Our analysis found 29 occurrences of this case (b in Table 3), representing 0.45% of the internal added triples (h in Table 3), 0.41% of both triples added (i), and less than 0.01% of the total of triples in DSj+1 (j). We show the results regarding the way modi cations a ecting an internal element of a triple in uence addition, removal or modi cation of links. Table 4 presents the results for the cases M:1, M:2, M:3 and M:4 (cf. Figure 2).
Case M:1. Results in Table 4 (a) show that 496.397 triples were modi ed. However, these modi cation did not produce association with the addition of links, sharing the same subject of the internal triple.
Case M:2. The second column of Table 4 shows that 245 internal triples that were modi ed caused the removal of links (b). This represents 0.04% (j) of the 496.397 modi ed internal triples (e), 197.58% (l) of the removed links (g), implying that links removed were associated to more than one internal triple, and 0.04% (k) of all the 528.275 modi ed triples (i).
Case M:3. As the third column of Table 4 shows, we found that 21.899 internal triples that were modi ed caused the modi cation of external triples (c). This represents 4.41% (j) of the 496.397 modi ed internal triples (d), 68.69% (l) of the modi ed links (g) and 4.14% (k) of all the 528.275 modi ed triples (i). Figure 5 shows an example of an internal triple with object that changed from `Fund' to `Credit', which caused an e ect of changing the external dataset connected by `exactMatch' from `USA National Agricultural Library LOD' dataset to the `Germany National Library LOD'9 dataset.
Case M:4 The fourth column of Table 4 shows the total number of times that modi cations in triples did not cause modi cation in links. Even if there is an evidence of change in predicate or object of the triple, no change was detected in the predicate or object of the external link. This is a case which deserves further investigation, because it can lead to links semantically inconsistent.
This investigation aimed to detect unexplored situations regarding the way RDF simple and complex changes in triples are associated with link change actions in LOD datasets. We found that it is possible to interrelate distinct change operations of RDF triples with the way links change. This is a key aspect to inform automatic techniques to update the link statements or to support the editors of controlled sources to update their models. Our original ndings are thus useful to the design of link evolution mechanisms, which might be essential to enable the management of the LOD evolution and insure data quality over time. Our investigation highlighted the potentialities of our ndings for managing the evolution of life science LOD.
We discovered that, for Agrovoc, simple changes as addition or removal of triples had a minor impact on the dataset, because these changes corresponded 1.33% of the computed changes. The analysis of the addition indicated that almost all the internal triples added had an external link added simultaneously with the same subject. The case representing the addition with internal triples with no link being added represented only 1% of the instances found. This implies that Agrovoc dataset clearly applies the concept of Linked Data, linking 99% of their newly added triples to an external dataset. The impact analysis of the exclusion change operation showed the opposite situation. Only 3.85% of the internal triples removed between the Agrovoc releases resulted in an exclusion of the associated links to an external dataset. The other 96.15% identi ed cases showed that if an internal triple is removed, the connected link remained untouched, generating a broken link.
Our results indicated that the biggest impact is caused by complex modi cations of internal triples, which responded in 98.67% of all the computed changes. The rst sub-case regarding modi cations was related to the addition of a link when an internal triple is modi ed. We did not nd an instance of this case, which demonstrates that in Agrovoc when there is a modi cation of a predicate or object of a triple, no link to a dataset is created.
The second sub-case was related to the removal, showing that there is a relation between modi ed triples leading to an exclusion of links. In particular, the modi cation of the predicate or the object of 245 triples, resulted in the removal of 124 links. Further investigations are required to comprehend if the removal of links happened with the in uence of an internal triple with the same subject.
The third sub-case of the e ects of modi ed RDF triples indicated that modi cation of triples can lead to modi cations in links sharing the same subject. The fourth sub-case concerns the most frequent one, in which the modi cation of triples lead to link unchanged. This case needs additional studies to further observed to which extend these unchanged links remained semantically inconsistent due to the modi cations of the associated RDF triples.
These results indicate the need to improve the LOD maintenance mechanisms. The Agrovoc LOD dataset10 was released connected to several external vocabularies (e.g., NALThesaurus, Chinese Agricultural Thesaurus and DBpedia). This allows the interchange of information across multiple systems, as well as human and machine interpretation of interconnected vocabulary data. However, the results obtained in this investigation suggest that such initiatives are a ected by link maintenance issues. Broken and outdated links may introduce inconsistency in the use of LOD datasets whereas the manually maintenance is a hard and costly task. This indicates the need for investigations of new mechanisms to support the maintenance of links.
A removal or update of internal triples generate the need for inspecting links aiming to evaluate whether they remain consistent or not. Usually, this demands manual work, which are subject to resource constrains and fails (as our results indicate). The implementation of computational mechanisms that support this task can be valuable for LOD maintainers. The implementation of these mechanisms relies on the research of advanced techniques that are able to suggest changes and to identify inconsistencies in established links. Machine learning techniques, for instance, can be explored to suggest maintenance actions based on the structure of the linked datasets and previously constructed training sets.
We plan to improve the algorithms implementing our analyses for comparing more than two versions of the same dataset. This might enable handling several releases of the same dataset and enriching the results of our analyses. In addition, we aim to address novel challenges concerning the e ects of class level changes on links. The study of these issues must pave the way for the de nition and development of mechanisms suited to the maintenance of links without requesting the reapplication of link discovery techniques. This must bene t the RDF dataset maintainers and underlying software application relying on links to favor their execution and e ectiveness. 6
Handling the impacts of RDF datasets evolution in established links is essential to maintain the consistency and bene ts of LOD over time. This is specially true and valuable for the life science RDF datasets, which are heavily interconnected. In this paper, we investigated the behaviour of links based on the evolution of RDF datasets. We conceived and developed analyses based on the computation of RDF dataset deltas to understand their in uence in the links of the dataset.
We found that complex changes (modi cation of triples) generate further impact in the links than simple changes, such as simple (atomic) addition and removal of triples. This result stands for a starting point in the investigation of a solution to preserve links automatically over time. Future work involves the de nition and implementation of additional analyses and their execution with larger datasets in other domains. We plan to investigate the correlation between changes in ontological level and their corresponding instances. We will investigate the elaboration of a link maintenance mechanism informed by the lessons learned in these analyses.
This work was nancially supported by the S~ao Paulo Research Foundation (FAPESP) (grants #2017/02325-5, #2018/08082-0, #2018/05357-8 and #2018/141997)11.