Embedding Naive Bayes Classification in a Functional and Object Oriented DBMS

This thesis introduces two implementations of Naïve Bayes classification for the functional and object oriented Database Management System Amos II. The first one is based on objects stored in memory. The second one deals with streamed JSON feeds. Both systems are written in the native query language of Amos, AmosQL. It allows them to be completely generic, modular and lightweight. All data structures involved in classification including the distribution estimation algorithms can be queried and manipulated. Several optimizations are presented. They allow efficient and accurate model computing. However, scoring remains to be accelerated. The system is demonstrated in an experimental context: classifying text feeds issued by a Web social network. Two tasks are considered: recognizing two basic emotions and keyword-filtering.