Découverte de motifs centrée sur l'utilisateur

Pattern mining is an enumeration technique used to discover knowledge from databases. This Habilitation thesis summarizes our main contributions regarding user-centric pattern mining. First, we introduce the pattern-oriented relational algebra (PORA), which is the formalism used throughout the thesis. We add a domain operator to the relational algebra to generate hypotheses about the data and a cover operator to compare the hypotheses to the data. Beyond the declaration of mining processes, this algebra makes it possible to reason on the queries to deduce properties or to optimize queries with rewriting rules. A second part presents our work where the user’s preferences guide the mining of patterns. In other words, pattern mining is seen as an optimization problem where only the best patterns in the sense of a preference relation are preserved. For this purpose, the cover operator is implemented to play the role of preference relation by comparing patterns two-by-two in order to retain the best ones. Finally, we are interested in model construction to improve the complementarity between mined patterns. The last part details our contributions where the method of analysis of the patterns guides their discovery. With this vision, the patterns are analyzed with a sharpness proportional to their interest. Rather than mining all the patterns, it is then sufficient to sample them with a probability proportional to their interest for presenting them to the user. We are extending PORA to reformulate the principle of pattern sampling algebraically. We show the interest of pattern sampling for the construction of anytime and interactive systems. Finally, a conclusion summarizes and discusses several research perspectives