Olivier Pivert

Vocabulary elicitation for informative descriptions of classes

Linguistic descriptions of numerical data using a vocabulary defined as linguistic variables are particularly useful to help a user understand the content of a dataset. When dealing with data structured with classes, the relevance of the linguistic descriptions strongly relies on the adequacy between the vocabulary and this data structure. This paper proposes a criterion to quantify this relevance, understood as informativeness and measured in terms of specificity. It then proposes various strategies to elicit appropriate fuzzy partitions to define the modalities of relevant linguistic variables and it experimentally examines their performance on artificial data sets.

An Approach Based on Predicate Correlation to the Reduction of Plethoric Answer Sets

Seeking data from large-scale databases often leads to a plethoric answer problem. A possible approach to reduce the set of retrieved items and to make it more manageable is to constrain the initial query with additional predicates. The approach presented in this paper relies on the identification of correlation links between predicates related to attributes of the relation of interest. Thus, the initial query is strengthened by additional predicates that are semantically close to the user-specified ones.

Fuzzy query by example

This paper describes Fuzzy Query By Example, an approach helping users retrieve data without any prior knowledge of the database schema or any formal querying language. The user is solicited to evaluate, in a binary way, pre-selected items of the database. We provide a characterization-based strategy that identifies the properties shared by the examples (resp. counter-examples) positively (resp. negatively) evaluated by the user. These properties are expressed using linguistic terms from a fuzzy vocabulary to ensure that the user has a good understanding of the inferred query.

On Making Skyline Queries Resistant to Outliers

This paper deals with the issue of retrieving the most preferred objects (in the sense of Skyline queries, i.e., of Pareto ordering) from a collection involving outliers. Indeed, many real-world datasets, for instance from ad sales websites, contain odd data and it is important to limit the impact of such odd data (outliers) on the result of skyline queries, and prevent them from hiding more interesting points. The approach we propose relies on the notion of fuzzy typicality and makes it possible to compute a graded skyline where each answer is associated with both a degree of membership to the skyline and a typicality degree. A GPU-based parallel implementation of the algorithm is described and experimental results are presented, which show the scalability of the approach.