Ankesh Khandelwal

• Isn't default assumption of deny_all contrary to the theme of web documents? Information on the web is created for consumption by all, even though the host may want to hide internal informations, and give restricted access to other sources.
• If we keep allow_all as default, most of the policies would be to deny access. In which case, by the current approach, the system would be post-filtering most of the time. How helpful would query expansion be in this scenario?

Comment: The authors have chosen to omit discussions on rho-iso and rho-join semantic associations. I believe that for each kind of association required different approach to semrank evaluation. For eg.

• in rho-iso association relation between rm and rn (last elements of the two paths) would be important. Association would be entirely different if rm and rn belong to disjoint classes, and not same class.
• Similarly, for rho-join association the relation between properties on the path from r1 to rn and those on the path from r2 to rn may matter. Again, association may be different if the two paths contain disjoint properties, rather than similar (including subProperty) properties.

1. I think the paper presents a great piece of work. The idea and results are superb. However, I am skeptical on the use of the word ontology. Does the word Ontology trouble you too? Semantic web is about linking resources and documents on the web with more than just hyperlinks. The paper presents techniques to relate concepts as close (similar) and generic-specifc (subclass). It creates a thesauri and extends it with subclass relationships. To me it can be called more than just a thesauri only if it contains relations other than similarity and subclass between its concepts.
2. A system that supports tagging can use the ontologies of concepts to classify things better (things tagged by similar terms/ concepts can be kept together). Which other semantic web application can use the ontologies of concepts?

1. In the paper under local approach weights are assigned to edge types. i.e. if nodes x and y are connected by edge label l irrespective of x and y this edge has same weight for all such x and y. My question is, that by this aren't we loosing some local information? Saying two FBI agents are friends has very different significance from two terrorists being friends and much more when a terrorist and FBI agent are friends (it may not be realistic, is used to only convey the idea).
2. Could loosing local information, mentioned above, be one of the reasons why local approach has not performed as well as global? For the given evaluation isn't it easy to come up with useful 'edge-label n-grams' or 'top edge-label n-grams' that could significantly boost precision after learning? i.e. Do you think that, since the relations that evaluation is looking for in the individual datasets are very few and narrow, it is easy to come up with few n-grams that can help improve the precision?

1. In section 3.2 authors simply mention 'we present the semantics for the notion of entailment for temporal graphs based on the corresponding notion of RDF graphs'. Later they use the notion of entailment in RDF graphs to prove Theorems 3 and later. I believe that it would have been convenient for readers if authors had included a line for RDF entailment (and-or a reference), like G1 entails G2 if an instance of G2 is subclass of cl(G1). What do you think? Have I missed something that authors have said in the paper?
2. I was expecting more from this paper. My feeling is that even with timestamping their work is scoped to snapshot. The temporally reified statements are used to group them into graphs with same time-stamps for all purposes (entailment), except for probably querying. They leave incorporation of <, >, = comparisons for future work. I was looking for entailments such (a, b, c){1,5} entails (a, b, c){4}, by inclusion of some axiomatic triples or other means.
3. There is another thing that may merit attention- transitivity of intervals. For example, (a,b,c){1,3}, (a,b,c){3,5} entails (a,b,c){1,5}.

• In section 5.1 the selectivity estimation for bound object is done separately for bound and unbound properties.
  • This is contrary to the assumption that selectivities of s, p , o are independent.
  • When p is unbound the selectivity value can be >1, not just theoretically.

• Selectivity Estimation is so naive, rudimentary. They should have at-least tried to estimate sel using sel(s)*sel(s/p)*sel(o/s,p) or sel(s)*sel(s/p)*sel(o/p) and suggested the required statistics and how these statistical informations be created. Moreover, this involves pre-processing (nothing done at query time). Only negative impact is the increase in size of summary data, but that's affordable.

1. In section 4.1 the paper says n-arity relationship should be reified as a group of binary relationships. I couldn't create a clear picture how this would be done, especially what would the namings of the relations be. Would it be the attribute name that is not part of primary key? What does reified mean here? How would tokens be affected?
2. In section 4.3, paper validates relationship between attribute id in author and hasID in ontology. I couldn't understand why would they be mapped in first place? Because from (2) VD(id) would contain Des(author) where as from (4) VD(hasID) would contain Des(Paper). What is confident mapping (good confidence measure)? Could you help with a better example for validation?
3. In section 4.4, the algorithm distinguishes categorical attribute to non-categorical. However, it isn't clear to me how do we determine if an attribute is categorical? A naive description can be any attribute not part of primary key is categorical. Can this be correct?

• How do we obtain L1 from L0?
• How is centroid c computed in step 9?

1. I wish authors could have written a line describing MI & PMI. It would be useful to extend this approach to consider semantic relations between labels (subPropertyOf).
2. I perceive a disconnect between method of finding outliers and the explanation that is generated. Do you see that as a problem? (Has it got anything to do with getOutliers(profile) algorithm?) Outliers are found using distance based algorithm, where as explanation is generated by selecting the features used by the classifier to classify the nodes into outliers and non-outliers found earlier. I'm not really able to connect the two things. Because o/w we can simply train our classifier to classify based on feature selection- and we can isolate outliers from it.

Casually: Can nSPARQL have a better representation? Isn't SPARQL (& SQL) very intuitive?

1. authors are working within the OWL Horst semantics and in section III-B they mention of Rule-base partitioning. I believe rules in OWL Horst are of the type {p a owl:TransitiveProperty, s1 p o1, o1 p o2 --> s1 p o2}. This rule can be re-written for each transitive property as {s1 tp o1, s2 tp o2 --> s1, tp, o2}, where tp is declared to be transitive in the KB. I couldn't clearly interpret the kind of rules the authors are referring. First of all does it matter? Although from the sole example they present it looks like they take the latter approach, and I believe rule base partitioning makes sense only in the latter approach. (Please correct me). Although latter approach means that rule base is comparatively huge because corresponding to each instance of the axiom they would have a rule.
2. I am concerned of the choice of data sets. First of all authors do not mention anything about the OWL expressivity of LUBM, UOBM or MDC (UOBM uses a more expressive OWL). Secondly, LUBM (similarly UOBM) is a benchmark for traditional OWL reasoning. By nature the individual university data sets are totally unrelated, and therefore speed-up on data partitioning is not a surprise. I do not mean to say that real life data sets are any different. Do you think that it would have been helpful if authors could have referred to some study that discussed characteristics of real-life rdf graphs (owl data) in terms of connectedness/ partitions etc. or at least give some sense of real rdf graphs?

1. In section 5.2, Property Characterisitcs, author chooses to restrict the range and domain of a symmetric property. (My feeling is that symmetric properties would most of the times have the same domain and range, and therefore the rule may not be useful). However my real question is- does this really capture the semantics of symmetric property. Do we not require similar rule as for inverse property. In this case it would be like If E has the property P1 whose value is V and the property P1 is 'symmetric' then V has P1 whose value is E. Or it is enough to say If the property P1 is 'symmetric' then P1 is 'inverse of' P2.
2. In section 5.3, Class Equivalence, author defines the rule If C1 is an equivalent class of C2 and E has C1 whose value is V then E has C1 whose value is V. Here, what does E has C1 mean?
3. This refers to the same quote as above. Why does author choose to show equivalence between classes only when that class is part of an equivalence axiom. i.e. author chooses to say, C1 equivalent to C1 if there-exists C2 such that C1 is equivalent to C2. why is there no need to declare C1 equivalent to C1, for rest of the classes?

1. In section 4.3.1 authors mention about distributive and collective reading as they choose the former, i.e. the certainty attached with G1 in RDF is considered to be certainty attached with each RDF statement in G1 and the certainty of all facts (n in number) in G1 considered together is certainty*certainty*....n-times. Have I understood it correctly? If not please do not read further! Could we considered certainty attached with G1 in terms of fuzzy logic rather than in probabilistic sense? Further, adopting the same distributive reading we could argue that the certainty of all facts in G1 considered together is minimum of certainty of all the facts, instead of their products?
2. There is a possibility that certain meta-property values are available for G1 and not for G2. Have authors discussed this possibility while querying named graphs G1 and G2? If I have missed something, again don't read further! Specifically consider Example 5.9. What happens if Iftime(x1) is known but Iftime(x2) is not known and we are considering Iftime(x1 intersection x2).
3. Comment: Idea of replicating meta statements on graphs to individual RDF statements doesn't appear neat! Lots of redundancy, unnecessary load on storage space.

1. Authors quote numbers from a reference to show that the number of OWL ontologies is 6 times lesser than RDFS ontologies. I bet there are very few OWL ontologies that use cardinality restrictions. Inspite of these statistics I am tempted to ask- wouldn't it be relevant to discuss cardinality restrictions in the context of the work in this paper. Cardinality restrictions (special case- functional properties) on a property, relate a class (RESTRICTION) to the range of the property. In addition they restrict the out-degree (total-degree) which is very relevant here. Do you think the same, and if possible how do you think cardinality restrictions can be accounted for?
2. This probably should have come before the previous question, but since first question led me to think of this one, i keep it second. Many properties are intended to take multiple values. For example a property that relates team to its members, say hasMember. If authors would have considered instance data this would have been accounted for. Therefore it may not be enough to consider a single edge <team, hasMember, players> for total-degree estimations. Do you also see that as a problem? How do you think we can resolve this issue, if there is one?
3. This question is also in context of those minority ontologies in OWL. I could understand that in the case of allValuesFrom, say C = AllValuesFrom(P,A), we can have an edge <C, P, A>. However, it's not clear from the paper that how the authors take into account someValuesFrom feature of OWL.
4. What about property hierarchy? It's an RDFS feature. My personal feeling is that its inclusion shouldn't affect the VR and CCDF values for total-degrees (do you think so?). Still.. do you think property hierarchies should have been discussed (or considered) by the authors?

• @Pg 10- Memory usage for data-sets of sizes 14, 255, 14, 556 are 20, 270, 10, 501 MBs respectively. In 2 cases there have been significant %age drop in storage sizes and in other 2 there have been increase in storage sizes. How can this be explained? (number of resources vs. number of triples for each of them?)
• Casual Question- The motivation is drawn from examples in Anti-money Laundering, Threat Assessment and Risk Assessment. However, the system has been tested only for Insider Threat Project, whose expected data set size is small. Aren't above data sets too huge, to keep memory based associations feasible?