We proposed a modular framework for RDF data exploration concretized in a software that encompasses all the research efforts to reach each objective.
Objective O1 is to extract meaningful knowledge from RDF graphs. This knowledge is at the basis of data exploration as it represents data in terms of their properties and avoids flooding users with too much punctual information. To achieve O1 we first gave a formal definition of the notion of knowledge and then provided ways to extract such knowledge directly from RDF data.
We defined knowledge in terms of “multi-dimensional aggregates” (MDAs). Consider a dataset about CEOs; an MDA, e.g. would be “Average age of CEOs grouped by nationality and number of managed companies”. The dataset itself contains all the data of each CEO (e.g. name, surname, age, nationality, etc.), the MDA, instead, exposes a more general property, that is, aggregated data.
Formally, an MDA is composed of: a topic of interest (the CEOs), the dimensions of analysis (nationality and number of managed companies), the measures (age) and an aggregation function (average). We provided a formal semantics of MDAs and an extensive set of options to identify them directly from RDF graphs. We also proposed ways to automatically derive, from the original data, novel dimensions and measures. This feature greatly enriches the pool of candidate MDAs, thus providing several new angles of analysis.
Objective O2 is to suggest interesting or unexpected aspects of the data. Several meaningful knowledge can be extracted from RDF graphs. For the CEOs, we extract the “Average age of CEOs grouped by nationality and number of managed companies” and also the “Sum of the net worth of CEOs with political connections, grouped by country of origin”, etc. O2 is to identify, among these MDAs, the most interesting or unexpected ones.
To achieve O2 we first defined what interesting or unexpected means when dealing with MDAs, then we provided several strategies to isolate the best ones. Aggregates whose result shows a trend or unexpected peaks are consider interesting. We quantify their interestingness using the statistical moments. Then we look for the MDAs with the best score.
W generate many candidate MDAs and determining their score ca be quite costly depending on the size of the data. Being able to do it in reasonable time is crucial, thus, efficiency played a key role in O2 and in the project as a whole. We proposed novel techniques to efficiently (i) quantify the interestingness of an MDA by simultaneously computing the results of several MDAs and (ii) prune uninteresting MDAs using early stopping to identify, as soon as possible, those that we can be sure (based on strong statistical evidence) are not the best ones.
Objective O3 is to compare knowledge extracted from different sources and grasp similarities or differences between them. A source about CEOs shows the “Average age of CEOs grouped by nationality and number of managed companies” whereas a source about Managers of organizations might show a different trend in the same knowledge. This insight shows that CEOs behave differently than managers.
RDF natively allows to model heterogeneous data. A resource in the data might be of type CEO while another might be of type Manager. Moreover, it allows to define hierarchies between types, e.g. we can state that both CEO and CTO are subtypes of Manager. Thus, all CEOs are Managers and all CTOs are Managers as well.
To reach O3 we take advantage of such hierarchies. We extract MDAs from resources of different types, then, we provide a visual comparison of the found MDAs. We compare MDAs with the same dimensions, measure and aggregation function. This comparison is performed among resources whose types are part of the same hierarchy. Thus, given the MDA “Average age grouped by nationality and number of managed companies”, we visualize its results considering the three heterogeneous sets of resources: CEOs, CTOs and Managers.
The results of our research have been published and presented on two international and one national conference. We took part to 4 poster sessions and disseminated our results through a website, regular seminars and social media activities.
