Most Linked Data browsers list all facts about an entity in an equal manner. In this paper, we present a prototype for identifying unexpected facts about entities, i.e., those facts that deviate from the expectations. To that end, we use an attribute-wise method for anomaly detection, which is also capable of providing qualitative explanations for the anomalies found. By comparing an entity at hand to a reference set of similar entities, we can provide information on how the entity at hand di ers from the typical patterns found for similar entities, and display those unexpected facts together with a short explanation.
Many linked data browsers display lists of facts about an entity at hand without
a particular notion of order or importance [
A slightly di erent problem is the search for unexpected or surprising facts. Rather than ranking facts by their importance, ranking by unexpectedness requires a notion of the usual state of an entity. To that end, an entity needs to be compared to a reference set of similar entities, and the typical patterns underlying the entities in that set have to be identi ed. Then, unexpected facts can be identi ed as those facts of an entity which strongly deviate from the patterns.
In this demonstration, we introduce a prototype for nding unexpected facts in DBpedia.1 Starting from selecting a DBpedia entity, the user can de ne the reference set and is then presented a number of unexpected facts. The rst results with selected entities look promising. The basic work ow of the tool comprises four steps, as depicted in Fig. 1. In the rst step, the user selects a DBpedia entity to analyze. This step is supported 1 Available online at http://topfacts.informatik.uni-mannheim.de/ 1
Lookup 2
categories 3
attribute-wise model learning 4
ABC verbalized rules by DBpedia Lookup and its autocomplete function.2 For example, the entity dbpedia:Bethlehem, Pennsylvania is selected.
Once the entity is selected, the user has to select a reference set of entities to
compare to. To that end, all YAGO types, which form a much richer hierarchy
than the DBpedia ontology types [
In the next step, the reference set is retrieved. For each entity, an attribute
vector is created, using attributes such as datatype properties and direct types.
Based on those feature vectors, an individual anomaly score for each attribute is
computed using the ALSO approach [
Finally, for all attributes that have a high anomaly score, the nding is
output as the models' justi cation for expecting a di erent value, ordered by the
respective anomaly score. An example output is shown in Fig. 2. Following the
details on demand paradigm [
Since the output would also show quite a few statements that are not unexpected facts, but mere errors in DBpedia, we lter out those rules referring to statements which are inconsistent with the DBpedia ontology.
For implementing the prototype, we use RapidMiner server5 with the Linked
Open Data extension [
In this section, we show some interesting example ndings for di erent resources and reference sets. 2 http://lookup.dbpedia.org 3 DBpedia delivers YAGO types as well, so no separate linkage to YAGO is required. 4 The numbers have been chosen for having a reference set that is big enough for discovering some meaningful patterns, and at the same time small enough to be processed in real time. 5 http://www.rapidminer.com Bethlehem, PA compared to Cities in Pennsylvania: Bethlehem is one of the oldest places in Pennsylvania, being founded in 1741. Furthermore, most cities in Pennsylvania are not a founding place of any organization, but Bethlehem is the founding place of Lehigh University Press.
Pennsylvania compared to States of the US: For Pennsylvania, we nd that it has some uncommon characteristics for the US founding states6: it is unusually large (119,283 square kilometers, with only New York being larger), has an unusually large maximum elevation (Mount Davis with 979m), and an unusually low area covered with water (2.7%).
Black Swan compared to Ballet Films: It is unusual, e.g., that Black Swan is an Academy Award winning ballet lm. Futhermore, ballet lms are usually not thrillers.
Trent Reznor compared to American Heavy Metal Singers: Unlike other heavy metal singers, Reznor is also a piano player and has written various lm scores. Joanne K. Rowling compared to British Billionaires: Rowling is the only female among the British billionaires, and one of the rare supporters of the Labour party. 4
In this paper, we have introduced a prototype which identi es unexpected facts about DBpedia entities. We compare an entity to a reference set of similar entities, and identify those facts which deviate from the patterns that are typical for the reference set.
While rst anecdotal ndings are promising, a full user evaluation, also contrasting di erent presentation variants, still has to be conducted. Such an evaluation should, ideally, try to de ne and capture the human notion of unexpectedness, which, however, is not trivial.
In our prototype, we have so far used direct types, numeric datatype
attributes, and relations as features. Other features, such as relations to
individuals or quali ed relations [
For de ning the reference set, we have only used YAGO categories so far. It would be interesting to also allow more sophisticated restrictions, e.g., compare a city to other cities in the same range of inhabitants.
In summary, the demo shows a novel way of interacting with Linked Data and identifying facts which are interesting to the user. 6 Although this was not the contrast set we chose, many of the rules found refer to the founding states.
Acknowledgements The work presented in this paper has been partly funded by the German Research Foundation (DFG) under grant number PA 2373/1-1 (Mine@LOD).