Baseline

Table 4: Baseline

’Baskarta - Vatten’ ’Nationalparker’ ’Fiskets geografier’

Matsvinn Luftkvalitet i Generellt

S¨odert¨alje strandskydd SCB - Milj¨o - Vattenkvalitetsdata Modellering av

Environment per grundvattenf¨orkomst V¨asternorrlands marina habitat och naturv¨arden Tr¨ad som f¨orvaltas Radonriskomr˚aden Mjukbottenfauna i

av tr¨ad och parker Ask¨o-Landsortsomr˚adet

Resultat av L¨ansstyrelsernas ˚Atg¨arder i

radonm¨atningar geodata vatten

Luftkvalitet och V¨anermuseet, Luftkvalitet och

meteorologiska data, objekt Enk¨opings

m˚anadsrapporter musem

Papperskorgar Armemuseet Teknikland

Osterg¨¨ otlands Geokritiska Ajtte, v¨axt

museum unders¨okningar

Insamling av avfall Nationella Utvidgat strandskydd

(Kolada) kalkdatabasen

Genomsnittlig k¨orstr¨acka Ostasiatisk¨ Bebyggelseregistret

med personbil, mil/ utst¨allning byggnad

personbil (Kolada)

Ekologiskt odlad ˚aker- Etnografiska F¨orvarsmusem

mark, andel (%) (Kolada) delobjekt Boden

30% SP 10% SP 0% SP

5 UP 8 UP 10 UP

These recommendations were randomly taken from the Environment theme, since each target dataset belongs to that theme. The results can be seen in Table 4.

9 Results

9.1.2 Algorithm 1: Structured information matching

Table 5: Structured information matching

’Baskarta - Vatten’ ’Nationalparker’ ’Fiskets geografier’

Skyddade omr˚aden,

Baskarta - H¨ojder tilltr¨adesf¨orbud Omr˚aden skyddade (f¨oreskriftsf¨orbud) enligt fiskvattendirektivet Skyddade omr˚aden, Milj¨o¨overvakningsstationer Bussh˚allplatser vattenskyddsomr˚aden f¨or s¨otvatten, kust och

- Ume˚a kommun hav samt badplatser Skyddade omr˚aden,

Grusytor f˚ageldirektivet Fiskf¨orekomster (Natura 2000, SPA) per l¨an

- V¨asterbotten Skyddade omr˚aden,

Papperskorgar naturreservat V¨ardefulla vatten - V¨asterbotten

Skyddade omr˚aden,

Cykelparkerings- djur- och v¨axtskydds- Limniska ekoregioner platser omr˚aden - Ume˚a kommun

Detaljplaner Tr¨ad som f¨orvaltas Skyddade omr˚aden, av gator och parker kulturreservat

H¨ogfluorerade ¨amnen (PFAS) och bek¨ampningsmedel:

Parkm¨obler Utvidgat strandskydd en sammantagen bild av f¨orekomsten i milj¨on.

Redovisning av ett regeringsuppdrag Tr¨ad som f¨orvaltas Radonriskomr˚aden Skyddade omr˚aden,

gator och parker nationalparker

Utvidgat Resultat fr˚an Gr¨on infrastruktur

-strandskydd radonm¨atningar analysresultat

Genomsnittlig k¨orstr¨acka

Radonriskomr˚aden med personbil, Nationella markt¨ackedata mil/personbil 2018; basskikt

(Kolada)

70% SP 100% SP 50% SP

2 UP 1 UP 2 UP

These recommendations were mostly sourced from the same publisher. The recommen-dations for Baskarta - Vatten does not seem to be of the same topic, but for the other two target datasets there are recommendations that seems to stick to a similar topic.

9 Results

9.1.3 Algorithm 2: LDA

Table 6: LDA, 100 topics

’Baskarta - Vatten’ ’Nationalparker’ ’Fiskets geografier’

Administrativ Liding¨os Evenemangs- En god v˚ard? ¨Overgripande indelning Inspire kalender indikatorer f¨or sjukv˚ardens

kvalitet och resultat Baskarta - Adresser Skolor via Karttj¨anst Meteorologiska observationer

- Radarkomposit Baskarta - Detaljer Statens J¨arnv¨agar Baskarta

- Ritningsregister FIRA

Descriptionis Ptolemaicæ Bygg och plantj¨ansten Baskarta - H¨ojder avgmentvm siue Occidentis - G¨allande planer

notitia breui commentario illustrata

Information om det G¨avles baskarta Kostnadsber¨akningar byggda kulturarvet,

nedladdladdningstj¨anst Andel milj¨obilar av totalt

Baskarta - V¨agar antal bilar i hela den Milj¨odataportalen geografiska kommunen,

(%) (Kolada)

Kommunala och Meteorologisk modell Bygga, bo och milj¨o

privata skolor KNEP - Liding¨o stad

Baskarta - Byggnader Beskattning, Uppgifter Detaljplaner och ˚atervinning f¨or inkomst˚ar f¨ore 2016 via Karttj¨anst

Kommunens skolor

med adresser K¨allor (¨oppna data) Grundvattenniv˚aer och koordinater

Soldata Naturtypskartering Meteorologiska observationer

(KNAS) - Radarbilder

80% SP 0% SP 0% SP

3 UP 9 UP 7 UP

These recommendations were sourced from almost only unique publishers for two of the target datasets. However, the recommendations seem to deviate a lot from the content of the target dataset.

9 Results

9.1.4 Algorithm 3: LSA

For this algorithm, two variations of LSA was tried out, differing only in the number of topics. In Table 7, 33 topics were chosen, while in the latter Table 8 100 topics were chosen.

Table 7: LSA 33 topics

’Baskarta - Vatten’ ’Nationalparker’ ’Fiskets geografier’

Skyddade omr˚aden, Luftpartikelm¨atningar Baskarta - H¨ojder vattenskyddsomr˚aden PM2.5 och PM10

- Ume˚a kommun

Skyddade omr˚aden, Milj¨o¨overvakning, havs-Baskarta - Detaljer tilltr¨adesf¨orbud och sj¨osediment

(f¨oreskriftsomr˚aden) (¨oppna data) - V¨asterbotten

Skyddade omr˚aden, Tillg¨anglighetsdatabasen Baskarta - Byggnader naturreservat & (TD)

- V¨asterbotten

Skyddade omr˚aden, L¨ansstyrelsers och Baskarta - Adresser djur- och v¨axtskydds- kommuners tillsynsverksamhet

omr˚aden - Ume˚a kommun enligt milj¨obalken f¨or 2015 Friluftsliv: Leder och Palsmyrar i Sverige G¨avles baskarta anordningar i skyddade 2008-2010

omr˚aden

Baskarta - V¨agar Riksintresse friluftsliv G ¨AU G¨ota ¨alvutredningen:

Erosion - datapaket Baskarta - Fastigheter Naturtypskartering G ¨AU G¨ota ¨alvutredningen:

(KNAS) Utf¨orda borrh˚al - datapaket Skyddade omr˚aden,

Tr¨adsk¨otsel f˚ageldirektivet Skolmat

(Natura 2000, SPA) - V¨asterbotten Livr¨addnings- PS.Skyddade omr˚aden,

utrustning Natura 2000, Ans¨okningsomg˚angar

visningstj¨anst.WMS

Markgeokemi, analysdata Luftkartl¨aggning Nationalparksplan fr˚an geokemiskt atlas

(¨oppna data)

100% SP 50% SP 0% SP

1 UP 2 UP 8 UP

For the Nationalparker target dataset, the recommendations seem to stick to the topic,

9 Results

while still finding some datasets from another publisher. For the other two target datasets, it is questionable to what degree the recommendations stick to a similar topic.

Table 8: LSA 100 topics

’Baskarta - Vatten’ ’Nationalparker’ ’Fiskets geografier’

Skyddade omr˚aden,

Baskarta - H¨ojder tilltr¨adesf¨orbud Fiskf¨orekomster -(f¨oreskriftsomr˚aden) WFS samt Shape

- V¨asterbotten

Skyddade omr˚aden, Fiskf¨orekomster - WMS Baskarta - Detaljer naturreservat

- V¨asterbotten

Skyddade omr˚aden, Biogeografiska regioner Baskarta - Byggnader vattenskyddsomr˚aden & - WMS

- Ume˚a kommun

Baskarta - Adresser Skyddade omr˚aden, Biogeografiska regioner djur- och v¨axtskydds- - WFS samt Shape omr˚aden - Ume˚a kommun

G¨avles baskarta AM.Skyddade omr˚aden Vattenf¨orvaltning - tilltr¨adesf¨orbud - WFS samt Shape Skyddade omr˚aden,

Baskarta - V¨agar f˚ageldirektivet V¨ardefulla vatten (Natura 2000, SPA) - WFS samt Shape

- V¨asterbotten

Baskarta - Fastigheter Skyddade omr˚aden, Vattenf¨orvaltning - WMS interimiska f¨orbud

Tr¨adsk¨otsel Utvidgat V¨ardefulla vatten - WMS strandskydd

Livr¨addnings Skyddade omr˚aden, Oceanografiska -utrustning biosf¨arsomr˚aden observationer

- Havsmilj¨odata Milj¨o¨ overvaknings-Luftkartl¨aggning F¨orbud mot stationer f¨or s¨otvatten,

markavvattning kust och hav samt badplatser

100% SP 60% SP 90% SP

1 UP 2 UP 2 UP

Increasing the number of generated topics results in recommendations that seem to have more similar topics to the target datasets, while decreasing the number of unique publishers.

9 Results

9.1.5 Algorithm 4: LDA trained on Wikipedia articles

Table 9: LDA trained on Wikipedia, 100 topics

’Baskarta - Vatten’ ’Nationalparker’ ’Fiskets geografier’

Parkm¨obler Giltiga ramavtal P˚ag˚aende sjukfall per diagnos, ˚alder 2005–

G¨asthamnar och J¨amf¨orelsedata F¨orskolor P˚ag˚aende sjukfall per diagnos, naturhamnar i Sk¨ovde kommun Sverige totalt 2005–

Meteorologiska Landskapsanalys av skogliga P˚ag˚aende sjukfall per observationer v¨ardek¨arnor i boreal region:

diagnos, l¨an 2005– - Lufttemperatur F¨orslag till skogliga v¨ardetrakter - arbetsmaterial

Kostnad per patient (KPP) Lst Riksintressen Cykelbanor - specialiserad psykiatrisk

¨

oppen- och slutenv˚ard Driftomr˚ade TN.Rail.DownloadService Flygskatt, skattesatser

f¨or flygresor

VV-Slitlager HH.EnvHealthDeterminant Statens Biografbyr˚a - register StatisticalData.Noise. ¨over videogramdistribut¨orer

DownloadService ¨over videogramdistribut¨orer Solelproduktion maj Platser och sev¨ardheter

2019 Region - Liding¨o stad Nyckeltal fr˚an Kolada Uppsalas solel

P˚ag˚aende sjukfall per

diagnos, specifikt f¨or TN.Rail.Nominal Omk¨orningsf¨orbud stressrelaterade sjuk- TrackGauge via Karttj¨anst

domar 2005–

Jvgdata UH Sjukfr˚anvarostatistik f¨or Nationella markt¨ackedata

kontraktsomr˚ade kommun- och 2018; basskikt

Bas regionanst¨allda

NVDB F¨orbud Laserdata Skog SMHIs portal f¨or

mot trafik ¨oppna data

20% SP 0% SP 0% SP

5 UP 9 UP 8 UP

The recommendations for this algorithm seem very random. One or two relevant rec-ommendations was found for each dataset. The diversity of publishers is high, with no recommendations from the same publisher as the target dataset for two of the target datasets.

9 Results

9.1.6 Algorithm 5: LSA trained on Wikipedia articles

Table 10: LSA trained on Wikipedia articles

’Baskarta - Vatten’ ’Nationalparker’ ’Fiskets geografier’

Baskarta - H¨ojder Information om det byggda Hammarkullen kulturarvet, hela Sverige i Minecraft G¨avles baskarta Skyddade omr˚aden kulturarv, Tillsynsverksamheter

enligt INSPIRE - Livsmedel

Asyls¨okande till Sverige Mobiln¨atens

Bilfl¨oden 1984-2017 t¨ackningsgrad

Skyddade omr˚aden

Potth˚al kulturarv, enligt Liding¨o Minecraft INSPIRE, hela

Sverige

Luftkartl¨aggning Historiska GIS-kartor NVDB Gatunamn Andel milj¨obilar av Cykelparkerings- V¨arldsarv i Sverige, totalt antal bilar i hela

platser enligt INSPIRE, den geografiska

nedladdningstj¨anst kommunen, (%) (Kolada) Detaljplaner V¨arldsarv i Sverige, L¨ansstyrelsernas

lokal-enligt INSPIRE, ytor och lokalkostnader visningstj¨anst

Landskapsanalys av Bullerutredning V¨arldsarv i Sverige skogliga v¨ardek¨arnor i

enligt INSPIRE boreal region: F¨orslag till skogliga v¨ arde-trakter - arbetsmaterial Hundrastg˚ardar Suecia antiqua Jvgdata R¨alsf¨orh¨ojning

et hodierna (tiff)

Fiskev˚ardsomr˚aden Luftpartikelm¨atningar Jvgdata

Trafikcentral-PM2.5 och PM10 omr˚ade

100% SP 0% SP 0% SP

1 UP 7 UP 7 UP

Some surprising recommendations were found in Table 10. For the target dataset Na-tionalparker, some interesting recommendations are found that are not found in recom-mendations from other algorithms. However, for Fiskets geografier target dataset there seem to be many unrelated recommendations. For two of the target datasets, none of the recommendations were from the same publisher.

9 Results

9.1.7 Algorithm 6: Doc2Vec trained on Wikipedia articles

Table 11: Doc2Vec trained on Wikipedia

’Baskarta - Vatten’ ’Nationalparker’ ’Fiskets geografier’

Skyddade omr˚aden, Fastighetstaxering,

G¨avles baskarta nautrreservat, typkoder

- V¨asterbotten

Skyddade omr˚aden, Beskattning. Uppgifter Baskarta - H¨ojder vattenskyddsomr˚aden f¨or inkomst˚ar f¨ore 2009

- Ume˚a kommun

BR.Gr¨ans f¨or skogsskydds-Baskarta - Adresser Utvidgat strandskydd best¨ammelser uppr¨attad

av Skogsstyrelsen Skyddade omr˚aden,

Baskarta - Detaljer tilltr¨adesf¨orbud (f¨oreskrifts- NVDB - Hastighetsgr¨ans omr˚aden) - V¨asterbotten

Baskarta - Byggnader PS.Biotiopskydd beslutade Exploateringskontorets av Skogsstyrelsen byggnadsregister Baskarta - V¨agar Skogliga v¨ardek¨arnor 2016 Beskattning.

Beskattnings-utfall (inkomst˚ar 2018) Platser och sev¨ardheter Samr˚ad om

Parkeringsautomater - Liding¨o stad uppgiftsinsamlingar/enk¨ater 2003-2018

US.AdministrativaOchSociala Bullerutredning TN.Rail.Nominal- OffentligaTj¨anster.Myndighets

TrackGauge OchKommunkontor.

Tillg¨anglighet K¨allor NVDB - B¨arighet

Biotopskyddsomr˚aden Beskattning, Debiterings-Hundrastg˚ardar beslutade av statistik (inkomst˚ar 2018)

Skogsstyrelsen

100% SP 40% SP 0% SP

1 UP 4 UP 6 UP

For the target dataset Nationalparker, the recommendations are mostly from the similar topics. For the target dataset Fiskets geografier, the recommendations does not seem to be related. Finally for the third target dataset Baskarta - Vatten the recommendations seem to be of a similar topic, however they are not very diverse with respect to publisher.

9 Results

9.1.8 Algorithm 7: Weighted Ensemble Recommender

Table 12: Weighted Ensemble Recommender 1:2

’Baskarta - Vatten’ ’Nationalparker’ ’Fiskets geografier’

Skyddade omr˚aden,

Landskapsbilds-Grusytor naturrreservat, skyddsomr˚ade

- V¨asterbotten

Papperskorgar Utvidgat strandskydd Radonm¨atning Skyddade omr˚aden,

Parkm¨obler tilltr¨adesf¨orbud (f¨oreskrifts- Baskkarta - Vatten omr˚aden) - V¨asterbotten

Skyddade omr˚aden, L¨ansstyrelsers och Tr¨adsk¨otsel f˚ageldirektivet (Natura 2000, kommuners

tillsyns-SPA) - V¨asterbotten verksamhet enligt milj¨obalken f¨or 2015

˚Atervinningscentraler Biogeografiska regioner

och ˚atervinnings- (Art- och habitatdirektivet) Skogliga v¨ardek¨arnor

stationer 2016

Geotekniska Friluftsliv: Leder och

unders¨okningar Skogliga v¨ardek¨arnor 2016 anordningar i skyddade

unders¨okningar omr˚aden

Radon Baskarta - Vatten Riksintresse friluftsliv Utsj¨obanksinventeringen Resultat fr˚an radon-, Riksintresse friluftsliv omg˚ang 2,

Artutbrednings-m¨atningar modeller: Argos Inre

och Yttre Grund Skyddade omr˚aden, Kartering av koninuitetsskog i Skyddade omr˚aden, vattenskyddsomr˚aden boreal region - arbetsmaterial vattenskyddsomr˚aden

- Ume˚a kommun - Ume˚a kommun

Fiskev˚ardsomr˚aden Stora skogs¨agares mark- VISS - Vatteninformations-innehav (2017) system Sverige

80% SP 40% SP 0% SP

2 UP 3 UP 6 UP

The recommendations seem overall to be of similar topics. Some diversity of the pub-lishers of the recommendations was noticeable.

9 Results

9.1.9 Gold Standard

Table 13: Gold Standard

’Baskarta - Vatten’ ’Nationalparker’ ’Fiskets geografier’

F¨orbud mot Skyddade omr˚aden,

markavvattning vattenskyddsomr˚aden Fiskev˚ardsomr˚aden Skogliga grunddata Skyddade omr˚aden ,

- Medelh¨ojd kulturarv enligt Vinterv¨agh˚allning INSPIRE

Fornl¨amningar och ¨ovriga Skyddade omr˚aden, Detaljplaner kulturhistoriska l¨amningar tilltr¨adesf¨orbud (f¨

oreskrifts-- nedladdningstj¨anst omr˚aden) - V¨asterbotten

˚Atg¨arder i vatten Kartverktyget Laddstationer Skyddad natur - G¨allande planer Skogliga grunddata Information om det

- Medeldiameter byggda kulturarvet, Bussh˚allplatser hela Sverige

Skogliga grunddata Skyddade omr˚aden,

- Grundyta biosf¨arsomr˚aden F¨ardtj¨anstzoner Klara dagar, januari Skyddade omr˚aden, Hydrologisk modell

och juli [serie] naturreservat - V¨asterbotten (S-HYPE2016) Klara dagar, ˚aret Bebyggelseregistret Potth˚al

anl¨aggning Oceanografisk

prognos-modell (NEMO) Bebyggelseregistret Papperskorgar - historiska modelldata byggnad

nedladdningstj¨anst

V¨ardefulla vatten - WMS Geodata Hundrastg˚ardar

0% SP 30% SP 0% SP

7 UP 4 UP 2 UP

The recommendations for two of the target dataset have the same publisher as the target dataset. The degree to which the recommendations are relevant seem to vary however.

9 Results

9.2 Representing recommendations

The top ten recommendations for each dataset are represented as a graph using the RDF data model.

Figure 22 An annotation that contains the target dataset and its top ten recommended datasets in RDF.

In Figure 22 the final result of the recommendations for one dataset is displayed. An annotation has the target dataset connected with the predicate oa:hasTarget. The recom-mended dataset are placed in the oa:hasbody predicate. An ordered collection contains ten ordered collection pages, which each has a dataset in oa:items. By using the as:first and as:next, the order of the recommendations is explicitly stated in the data.

10 Discussion

10 Discussion

This project has shown that it is indeed possible to infer relations between datasets.

However, a challenge has been to find an evaluation method to accurately assess the quality of the inferred relations. While metrics such as Unique Publisher was used to gain a sense of diversity among the created recommendations, diversity in itself may not provide a good assessment of quality, as its relevance may vary depending on the circum-stance . For some of the datasets, the best recommendations might reside in the datasets published by the same publisher as the target datasets, while in other cases it may not.

However, there seems to be some relevance to some of the created recommendations, especially for the weighted ensemble model, that was considered to perform best. For example, the middle column in Table 12 does show that some recommendations does indeed seem relevant. But overall, the quality of the created recommendations varies, with some seemingly random recommendations.

The algorithm that matched the structured information of datasets resulted in some interesting recommendations, as can be seen in Table 5. However, diversity, or variance, of the publishers of the recommendations was not very high. Other approaches were therefore taken in which the free text metadata was utilized. This was investigated by using the LDA and LSA feature extraction methods, which resulted in more diversi-ty/variance in terms of the publishers of the recommendations. The stochastic LDA method, in particular, resulted in recommendations with high diversity, however the recommendations seemingly did not stick to the topics of the target dataset. Instead, the LSA method resulted in recommendations that seemed to be more relevant to the target dataset, especially for the LSA method using 100 topics instead of 33 topics. With a higher number of topics, the result seems to be that datasets with similar content to that of the target dataset are found. However as the number of topics increased, the diversity, or variance, decreased.

In order to retain a high degree of diversity, while still generating relevant recommenda-tions, transfer learning was explored for the LSA, LDA and Doc2Vec feature extraction methods. Transfer learning of the LDA model had seemingly little effect, neither on the topics of the recommendations or the degree of diversity. For the LSA method, on the other hand, there was a substantial increase in diversity, but once again, the topics of the recommendations deviated at times from the target dataset. For the Doc2Vec method, the diversity was moderate, with some interesting recommendations but also seemingly random ones. Finally, an ensemble recommender system was tested using the pre-trained Doc2Vec model and the theme metadata. The idea was to mix the two algorithms to reduce the variance, while minimizing the effect of the potential bias in the data. This approach yielded interesting results. The algorithms that we believe to have performed best was Algorithm 7.4 based on LSA with 100 topics and Algorithm 7.8, using the ensemble recommender system approach. For two of the target datasets, Nationalparker and Fiskets geografier, these two algorithms matched, or even surpassed, the

recommen-10 Discussion

dations of the gold standard algorithm. For the third target dataset, Baskarta - Vatten, none of the created algorithms were capable of surpassing the gold standard in terms of finding related datasets. A brief discussion of the possible reasons for the outcomes of the discussed algorithms are presented below.

Worth mentioning is that the created models themselves may not be entirely to blame for the varying quality of the recommendations. If one observes the data that is used to create recommendations, one can observe what seems to be interesting (and perhaps not totally unexpected patterns). If a dataset has a low quantity of text metadata, as is the case for the target dataset Fiskets geografier, the model seems to have a hard time finding similar datasets, as can be seen in the right column of Table 12. However, as can be seen in the recommendations for the target dataset Baskarta - Vatten (left column of Table 12), raw quantity of is not enough either.

To further accentuate the influence that the metadata seems to have on the created re-lations, we can once again look at the middle column of Table 12, whose corresponding target dataset Nationalparker possess some degree of quality and quantity. Given text of both quality and quantity, the model seems capable enough to actually create relevant recommendations. While more thorough evaluation of the quality of the recommenda-tions is necessary, the quality of the created recommendarecommenda-tions appear to be tied to the quality of the metadata, which is not totally unexpected.

Another aspect related to the overall quality of the created algorithms is bias. As pointed out in the Section 5, at the time of writing, a small number of publishers accounted for a large proportion of all published datasets. The elements consisting of free text were completely provided by the publisher, meaning that the publisher had total freedom when writing the text. There were cases where the author-provided free texts for descriptions and keywords were very similar for datasets published by the same publisher. One such example was the target dataset Baskarta - Vatten that was used in the evaluation. Phrases and pieces of text were re-used for multiple datasets. Similarly, publishers often included their name in the metadata. Thus, patterns arose in the metadata that did not reflect the actual contents of the datasets. Several approaches to mitigating the effects of this bias were investigated. For instance, removing the name of the publisher from the textual data, utilizing transfer learning on the Wikipedia dataset, and increasing the variance of the recommendations in terms of publisher. The idea was, by performing these additional steps, an algorithm could be created that was less prone to be affected by unwanted patterns in the data. As previously discussed, trying to reduce the bias often resulted in a higher variance of the publishers. The ensemble recommender system approach proved to be the most effective at reducing the variance, while also minimizing the effect of bias in the data

While it is difficult to know how well the created algorithms perform, we believe it is safe to argue that there is much improvement to be done in terms of cultivating the quality of future metadata. As more datasets are added, a more even spread of datasets can very well be achieved, which would counteract some of the bias.

10 Discussion

We believe that it is likely that the number of datasets will increase over time. As mentioned in introduction, the number of published datasets have tripled over the course of this project. Moreover, not all organisations have yet published metadata about their datasets on the dataportal. For instance, there are 290 municipalities in Sweden [skr20]. At the time of writing, only thirteen municipalities have published metadata

We believe that it is likely that the number of datasets will increase over time. As mentioned in introduction, the number of published datasets have tripled over the course of this project. Moreover, not all organisations have yet published metadata about their datasets on the dataportal. For instance, there are 290 municipalities in Sweden [skr20]. At the time of writing, only thirteen municipalities have published metadata

In document Inferring Dataset Relations using Knowledge Graph Metadata (Page 49-71)