Inductive learning of OWL 2 property chains

Tracking #: 2235-3448

Jedrzej Potoniec

Responsible editor: 
Jens Lehmann

Submission type: 
Full Paper
We present an algorithm to inductively learn OWL 2 property chains to be used in object subproperty axioms. For efficiency, it uses specialized encodings and data structures based on hash-maps and sparse matrices. The algorithm is based on the frequent pattern search principles and uses a novel measure called s-support. We prove soundness and termination of the algorithm, and report on evaluation where we mine axioms from DBpedia 2016-10. We extensively discuss the 36 mined axioms and conclude that 30 (83%) of them are correct and could be added to the ontology.
Full PDF Version: 


Solicited Reviews:
Click to Expand/Collapse
Review #1
Anonymous submitted on 09/Jun/2019
Review Comment:

This manuscript was submitted as 'full paper' and should be reviewed along the usual dimensions for research contributions which include (1) originality, (2) significance of the results, and (3) quality of writing.

The paper proposes an approach for the extraction of OWL2 sub-property chains and derive axioms for Web of Data knowledge bases.
The author evaluates the approach, based on frequent-pattern/association rules mining, on DBPedia and discusses the effectiveness of the approach case-by-case.

The solution may be potentially interesting for the Semantic Web/Web of Data community. However, I think that the paper is too preliminary for the acceptance as it is
There are several problems to be addressed regarding the positioning w.r.t. other existing approaches, the kind of approach, its generality and the evaluation. All these aspects are below the standard for a journal publication.

As regards the positioning of this work, it is not clear the need for this approach. I agree that a well-known feature of OWL2 language is the construction of property chains but the authors did not clarify it in the paper. Which are the advantages w.r.t. the existing approaches?
The framework is another approach based on association rule mining. This makes the system similar, in the spirit,
to other association rule miners for the Web of Data but they are not mentioned in the paper. For instance systems such as AMIE(++)[1,2]/CARL[3] and those proposed by d'Amato et al.[4,5] are not cited.
In these works, the goal was to find (SWRL)rules from an ontological knowledge base.
Although these systems are used for predictive purposes, I believe that the systems may be easily adapted to support descriptive tasks or, at least, the property chain miner proposed in this submission
may be used for prediction (e.g. using only high-confidence rules).
Closer works that are not mentioned are those proposed by Sazonau et al.[6,7] where the terminology of the knowledge base was induced using association rule mining methods (DL-Miner) based on a refinement operator.
Less recent works on the extraction of multi-relational rules were also proposed by Lawrynowicz et al.[8].

The discussion w.r.t. the other methods (like those reported above) should be provided extending the introduction and the related work section.
In the discussion, the author should answer to the following questions:
1- Is the system a complementary approach to the existing ones?
2- Which benefits can be obtained using this approach instead of another rule miner? (In this case, one should check if the rule miner induces also rules expressing sub-properties chains)

Moreover, I think that the paper should discuss better the measures introduced in Sect. 5: where is the novelty w.r.t. other measure adopted for frequent patterns/association rules?
Note that the measures proposed by Sazonau et al.[6] may be easily adapted to evaluate the quality of sub-property chain axioms.
Concerning the class expression learning problem (mentioned in the Related work), there are several other (recently) published methods
proposed in the literature in addition to DL-Learner (even if it is certainly one of the most famous): PARCEL[9], SPACEL[10], DL-Foil[11], DL-Focl[12], etc.

As regards the technical aspects of the proposed approach, the mining technique based on patterns/association rules elicitation proposed in this paper is sound but not complete, as stated by the authors in Sect 5.5.
This means that the algorithm may not terminate.
This is a not totally new consideration if one considers the multi-relational nature of the representation language (see references about WARMR and WARMeR[13] that are systems where the downward closure property w.r.t the support could not be satisfied).
This leads to a problem of efficiency that must be discussed (even in the evaluation when the maximum length parameter is set).

The experimental session is a problem with the highest severity. The deep insight of the results is interesting but it does not allow to get general conclusions about the generality and, importantly, the weakness of the approach.
In this perspective, considering only a KB represents another limitation of the paper. Thus I suggest an extension of the evaluation with more KBs keeping a similar insight (if possible).
Also, despite the fact that the author proposed a matrix based-encoding of the rules for efficient counting, the efficiency is not discussed in the paper.
The efficiency of any frequent pattern/association rule miner depends on the support estimation and the length of the pattern. However, the settings of the proposed approach is a little trivial, with a max_length=2.
Such aspects must be discussed because the support is affected by the sparseness of the assertions of the KB.
To this purpose, the author must stress the approach considering different min_supp and max_length values.
A further issue is the interestingness measure. In association rule mining, determining the interestingness of a rule is an open issue since it is a subjective criterion. This explains why there are several measures in the literature (e.g. lift, etc.).
Consequently, only reporting the adopted interesting measure (confidence) does not give insight into the effectiveness of the approach.
An experimental setting similar to the one adopted in [1,2] may provide a more objective viewpoint of the quality of the rules/chain properties.
Finally, the improved related work section may help to extend the evaluation comparing the outcomes w.r.t. a competitor.

[1]AMIE: association rule mining under incomplete evidence in ontological knowledge bases. WWW 2013
[2]Luis Galárraga, Christina Teflioudi, Katja Hose, Fabian M. Suchanek: Fast rule mining in ontological knowledge bases with AMIE+. VLDB J. 24(6): 707-730 (2015)
[3]Thomas Pellissier Tanon, Daria Stepanova, Simon Razniewski, Paramita Mirza, Gerhard Weikum: Completeness-Aware Rule Learning from Knowledge Graphs. International Semantic Web Conference(1) 2017: 507-525
[4]Claudia d'Amato, Steffen Staab, Andrea G. B. Tettamanzi, Duc Minh Tran, Fabien L. Gandon: Ontology enrichment by discovering multi-relational association rules from ontological knowledge bases. SAC 2016: 333-338
[5]Claudia d'Amato, Andrea G. B. Tettamanzi, Duc Minh Tran: Evolutionary Discovery of Multi-relational Association Rules from Ontological Knowledge Bases. EKAW 2016: 113-128
[6]Viachaslau Sazonau, Uli Sattler: Mining Hypotheses from Data in OWL: Advanced Evaluation and Complete Construction. International Semantic Web Conference (1) 2017: 577-593
[7]Viachaslau Sazonau, Uli Sattler, Gavin Brown: General Terminology Induction in OWL. International Semantic Web Conference (1) 2015: 533-550
[8] Joanna Józefowska, Agnieszka Lawrynowicz, Tomasz Lukaszewski: Towards Discovery of Frequent Patterns in Description Logics with Rules. RuleML 2005: 84-97
[9]Tran, A.C., Dietrich, J., Guesgen, H.W., Marsland, S.: An approach to parallel class expression learning. RuleML 2012. LNCS, vol. 7438, pp. 302–316. Springer, Heidelberg (2012)
[10]Tran, A.C., Dietrich, J., Guesgen, H.W., Marsland, S.: Parallel symmetric class expression learning. Journal of Machine Learning Research. 18, 64:1–64:34 (2017)
[11]Nicola Fanizzi, Giuseppe Rizzo, Claudia d'Amato, Floriana Esposito: DLFoil: Class Expression Learning Revisited. EKAW 2018: 98-113
[12]Giuseppe Rizzo, Nicola Fanizzi, Claudia d'Amato, Floriana Esposito: A Framework for Tackling Myopia in Concept Learning on the Web of Data. EKAW 2018: 338-354
[13] Bart Goethals, Jan Van den Bussche: Relational Association Rules: Getting WARMeR. Pattern Detection and Discovery 2002: 125-139

Review #2
By Md Kamruzzaman Sarker submitted on 21/Jul/2019
Major Revision
Review Comment:

This paper uses frequent pattern mining principles to find property chains in an ontology. It proposes a multiplication based algorithm to find the property chains. Along with the algorithm it represents some definitions and theorems required to prove the soundness, completeness, and termination of the algorithm. Evaluation of the algorithm was done using dbpedia ontology.

This paper focuses on one of the important problems of semantic web, the knowledge acquisition bottleneck. It proposes an algorithm to find property chains (new subproperty axioms from a given ontology), which can be added to the ontology after verification.

While frequent pattern mining is well studied in the field of data mining, it is not well studied in the field of semantic web. Frequent pattern mining is worth investigating for semantic web and authors did pretty well to investigate this. Although the algorithm is simple it gives options to find new axioms inductively. Authors also mentioned some technical details (using efficient encoding for search and query). They also give definition of support, s-support in terms of ontology. They provided some theorems to support the proposed algorithm.

The evaluation is done on dbpedia ontology, which is good, but as the algorithm is inductive, more experiments should have been done using different ontology evaluate the proposed algorithm properly. It is worth to see how it performs while we have a different kind of scenarios, for example when there is a very large number of predicates, but only small number of subject and object, or vice versa. Its worth to see how the performance varies. Another important aspect is the analysis of the complexity of the algorithm is missing.

In terms of originality, this paper introduces a technique, which was not previously studied for property chains learning in semantic web.

Outcome result from this algorithm is significant for the semantic web, but the authors mention a small amount of experiment. More experiment is needed to validate the algorithm.

Write-up is good, with some minor typos (PSO instead of POS in the last paragraph of section 4.2).

Overall this paper offers the use of an inductive algorithm to find object property chain axioms and evaluated the algorithm. Although theoretical analysis is clear, evaluation is only done on dbpedia ontology. This substantially decreases the value of the paper. As the algorithm is inductive, it is important to see how this algorithm performs in a different ontology.

Review #3
Anonymous submitted on 26/Jul/2019
Major Revision
Review Comment:

This is a review of the paper titled "Inductive learning of OWL 2
property chains."

Overall, we found the paper easy to follow and written in a clear
manner. We thank the author for their contribution.

The question of property chains (or property paths) warrants
additional investigation and research.

We believe that the following points should be addressed. Otherwise,
it might be easier to accept this contribution in some kind of workshop

* The comparison to general rule mining tools such as AMIE+ should be

* The contribution goes very much into detail about representing RDF
data in a Subject-Object matrix per property (Section 4). We believe
that this could be shortened. At the very least, the stipulated
performance improvements should be verified against some other
approaches (e.g. SPARQL query)

* The paper introduces a new s-support measure. The effectivity of
this measure should be compared. While the bounds are clear, it is
not obvious if there is any gain in doing the experiments with
s-support, limited s-support or just a regular support definition.

* It is possible to express the research question using an equivalent
SPARQL query. This should be compared and argued about.

* With regards to the evaluation, the paper title lays focus on
property chains. Thus, it would seem valid to exclude the property
chains of length 1 from the Table (Table 1) in the evaluation, and
instead present some propery paths of length > 2 instead.

* The system is evaluated on the DBpedia Mapping-based infobox
properties. It should also be evaluated on a further data set, for
example Wikidata.

* It would be nice to know run-time measures.