information theoritic analysis of entity dynamics on the linked open data cloud

www.moving-project.eu

TraininG towards a society of data-saVvy inforMation prOfessionals to enable open leadership INnovation

Chifumi Nishioka and Ansgar Scherp

ZBW -- Leibniz Information Centre for Economics and Kiel University, Germany

Information-theoretic Analysis of Entity Dynamics on the Linked Open Data cloud


2 of 19

Motivation

• Understanding the dynamics of the LOD cloud is important for many applications • e.g., SPARQL query caching, crawling strategies, term

recommendations

• Related work • Evolution of LOD documents [Käfer et al. 13] • Dynamics of LOD sources [Dividino et al. 14]

• Entities on the LOD cloud

• Used by a lot of applications • Knowledge graph in search engines • Document modeling [Schuhmacher and Ponzetto 14]

Chifumi Nishioka ([email protected])

Come to the presentation of

“TermPicker” by Johann Schaible at

14:30 on 1st June (Wednesday)

We conduct an analysis focusing on entities


3 of 19

Research Goals

• Measure the changes in entities between two points in time • Represent the temporal dynamics of entities as time-series

• Time-series clustering • Periodicity detection

• Evaluate four different features of entities


Goal 1: Represent the temporal dynamics of entities

Goal 2: Find out the representative temporal patterns of entity dynamics

Goal 3: Find out which features of entity more likely define temporal dynamics of entities


4 of 19

Formalization

• 𝑋𝑡: snapshot of LOD documents at point in time 𝑡 • Snapshot is a collection of triples 𝑥

• 𝑥: triple • 𝑥 = 𝑠, 𝑝, 𝑜 : subject, predicate, and object



5 of 19

Entity and Entity Representations

• Entities are represented by a set of triples • Entity Representation: Out

• Set of triples with common subject URI • e.g., db:John_Brown is defined by two triples

• Entity Representation: InOut • Set of triples with common subject URI or object URI • e.g., db:John_Brown is defined by three triples


db:Anne_Smith db:spouseOf db:John_Brown

db:John_Brown db:birthplace db:Los_Angels

db:John_Brown db:works db:Green_University


6 of 19

Triple Weighting

• example: Barack Obama

• <Barack_Obama, dbp:vicePresident , Joe_Biden> is more important than <Barack_Obama, rdf:type , foaf:Person>

• Baseline • All triples have a same weight

• Combined Information Content (combIC) [Schuhmacher and Ponzetto 14] • 𝐼𝐶 𝑣 = −log(𝑃(𝑣)) • 𝑝𝑟𝑒𝑑 𝑥 , 𝑜𝑏𝑗 𝑥 returns predicate and object of a

triple 𝑥, respectively


𝑤𝑏𝑎𝑠𝑒𝑙𝑖𝑛𝑒(𝑥) = 1

Each triple in entities has different importance for entities

𝑤𝑐𝑜𝑚𝑏𝐼𝐶 𝑥 = 𝐼𝐶 𝑝𝑟𝑒𝑑 𝑥 + 𝐼𝐶(𝑜𝑏𝑗(𝑥))


7 of 19

Measuring Entity Dynamics

• Cosine distance

• Euclidean distance


Goal 1: Represent the temporal dynamics of entities

𝛿𝑐𝑜𝑠𝑑 𝐸𝑡1 , 𝐸𝑡2 = 1 −𝐸𝑡1 ∙ 𝐸𝑡2

||𝐸𝑡1|| ∙ | 𝐸𝑡2 |

1. Measure the amount of changes in entities between two successive snapshots by one of two distance measures

𝛿𝑒𝑢𝑐 𝐸𝑡1 , 𝐸𝑡2 = (𝐸𝑡1,𝑖 − 𝐸𝑡2,𝑖)2

𝑖=1


8 of 19

Vector Representation of Entities

• Represent an entity 𝐸 by one-hot encoding • Extract all unique triples from different snapshots • Fix order of triples • e.g., db:Anne_Smith at 𝑡1 is (1,1,1,0,0) and at 𝑡2 is

(1,0,1,1,1)

• Cosine distance: 𝛿𝑐𝑜𝑠𝑑 𝐸𝑡1 , 𝐸𝑡2 = 1 −2

3∙ 4= 0.42

• Euclidean distance: 𝛿𝑒𝑢𝑐 𝐸𝑡1 , 𝐸𝑡2 = 3 Chifumi Nishioka ([email protected])

𝑡1 1 db:Anne_Smith db:birthplace db:New_York

2 db:Anne_Smith db:works db:Green_University

3 db:Anne_Smith db:spouseOf db:John_Brown

𝑡2 1 db:Anne_Smith db:birthplace db:New_York

4 db:Anne_Smith db:works db:Royal_University

3 db:Anne_Smith db:spouseOf db:John_Brown

5 db:Anne_Smith db:degree db:Master_of_Science


9 of 19

Temporal Dynamics of Entities

• Temporal Dynamics of an entity 𝐸

• 𝑛: the number of snapshots


Δ(𝐸) = (𝛿 𝐸𝑡1 , 𝐸𝑡2 , 𝛿 𝐸𝑡2 , 𝐸𝑡3 , ⋯ 𝛿(𝐸𝑡𝑛−1 , 𝐸𝑡𝑛))

2. Represent temporal dynamics of entities by a time-series of the amount of changes in an entity between two successive snapshots

Subsequently, we mine the resulted time-series to find out patterns of temporal dynamics of entities


10 of 19

Time-series Clustering


• Clustering algorithm: k-means++ [Arthur and Vassilvitskii 07] • Introduce an improved initial seeding into k-means

• Distance measure: Euclidean distance • The most efficient measure for distance between

time-series with a reasonably high accuracy [Wang et al. 13]

• Optimization of the number of clusters : Average Silhouette

Goal 2: Find out the representative temporal patterns of entity dynamics


11 of 19

Periodicity Detection

• Periodicity Detection

• A task of detecting periodicity from time-series • Example 1: (1, 3, 2, 1, 3, 2) -> periodicity of three • Example 2: (1, 2, 1, 2, 1, 2) -> periodicity of two • Employ a convolution-based algorithm [Elfeky et al.

05]


We assume that the amount of changes of entities have some periodicity

We see the centroids of the resulted clusters as patterns of entity dynamics


12 of 19

Dataset

• Dynamic Linked Data Observatory (DyLDO) dataset [Käfer et al. 12] • Weekly snapshots of the fixed set of LOD documents • 165 snapshots over three years (05/2012 to 07/2015)

• Entities in the DyLDO dataset • Almost 75% of entities appear only at one snapshot • Focus on entities that appear at >70% of snapshots


Entity

representation

# of unique

entities in 165

snapshots

# of entities that

appear at >70% of

snapshots

Out 27,788,902 2,909,700

InOut 29,097,929 2,950,533


13 of 19

Patterns of Entity Dynamics (1/3)

• Analysis with respect to eight conditions • Conditions are made by two entity representations,

two distance measures, two triple weighting methods

• Result of clustering • # of clusters are smaller when using combIC



14 of 19



Out Cosine Baseline Out Cosine CombIC

Out Euclidean CombIC

Out Euclidean Baseline


15 of 19



InOut Cosine Baseline InOut Cosine CombIC

InOut Euclidean Baseline InOut Euclidean CombIC


16 of 19

Periodicity of Entity Dynamics


We observe periodicities in temporal dynamics of entities

• e.g., “Periodicity of 56” indicates that the amount of entity changes vary along with one-year cycle

• Different patterns have different periodicities


17 of 19

Features for Entity Dynamics (1/2)

• Four features of entities • RDF Type (𝑓1) • Property (𝑓2) • Union of RDF types and properties (𝑓3) • Pay level domain (PLD) of entity URI (𝑓4)

• e.g., http://dbpedia.org/resource/The_Beatles -> dbpedia.org

• Evaluate four features by RandIndex • RandIndex: a metric of clustering • Measure the difference of clustering by a feature and

by entity dynamics (i.e., time-series vectors)




18 of 19

Features for Entity Dynamics (2/2)

• Entities that share a common PLD are more likely

to have similar temporal dynamics of entities when employing baseline for triple weighting

• When using combIC, entities that have a common RDF type or ECS more likely to belong a same cluster



19 of 19

Thank you for your attention!

Project consortium and funding agency


MOVING is funded by the EU Horizon 2020 Programme under the project number INSO-4-2015: 693092


20 of 19

Conclusion

• Temporal dynamics of entities on the LOD cloud • Represent the temporal dynamics of entities as time-

series • Find out the representative temporal patterns of

entity dynamics • Find out which features of entity

• Future work • e.g., SPARQL query caching




21 of 19

Appendix



22 of 19

Reference

• [Arthur and Vassilvitskii 07] D. Arthur and S. Vassilvitskii. k-means++: The advantages of careful seeding. SODA, 2007.

• [Elfeky et al. 05] M.G. Elfeky, W.G. Aref, and A.K. Elmagarmid. Periodicity detection in time series databases. IEEE TKDE, 2005.

• [Käfer et al. 12] T. Käfer, J. Umbrich, A. Hogan, and A. Polleres. Towards a dynamic linked data observatory. LDOW, 2012.

• [Käfer et al. 13] T. Käfer, A. Abdelrahman, J. Umbrich, P. O’Byrne, and A. Hogan. Observing linked data dynamics. ESWC, 2013.

• [Neumann and Moerkotte 11] T. Neumann and G. Moerkotte. Characteristic sets: Accurate cardinality estimation for RDF queries with multiple joins. ICDE, 2011.

• [Schuhmacher and Ponzetto 14] M. Schuhmacher and S.P. Ponzetto. Knowledge-based graph document modeling. WSDM, 2014.

• [Wang et al. 13] X. Wang, A. Mueen, H. Ding, G. Trajcevski, P. Scheuermann, and E. Keogh. Experimental comparison of representation methods and distance measures for time series data. Data Mining and Knowledge Discovery, 2013.

• [Yang and Leskovec 11] J. Yang and J. Leskovec. Patterns of temporal variation in online media. WSDM, 2011.



23 of 19

Entities in the DyLDO dataset

• Distribution of # of times of appearances of entities in 165 snapshots


Entity representation Out Entity representation InOut


24 of 19

Resulted Temporal Patterns (1/4)


Out Cosine Baseline Out Cosine CombIC


25 of 19



Out Euclidean Baseline Out Euclidean CombIC


26 of 19



InOut Cosine Baseline InOut Cosine CombIC


27 of 19



InOut Euclidean Baseline InOut Euclidean CombIC


28 of 19

Presentation

• PROFILES 2016


information theoritic analysis of entity dynamics on the linked open data cloud

Technology