Models and algorithms for recommendation and preference management in product configuration – BR4CP (Business Recommendation for Confi

The scientific and technical program will be organised into five main tasks:
* An upstream task of requirement analysis and definition of the benchmarks and of the protocol that will be used for validating the algorithms. The requirement analysis will include a detailed analysis of existing facilities offered by B2C recommender systems
* A series of studies developing preference learning models and algorithms for configuration systems, by applying concepts and techniques coming from machine learning.
* A third task aiming at designing search algorithms that will be efficient enough to be used online (e.g. SAT-based search, preference propagation).
A ``compilation of combinatorial models« task . The idea is to translate the obtained model and the recommendation model in a way that enables a quick online usage, for both optimisation and interactive resolution techniques.
* A last task devoted to an experimental comparison of the approaches advocated by the previous tasks

-----------

-----------

-----------

E-commerce, like classical commerce, adresses the problem of conciliating the different needs of customers, and the choices or even the purposes of the supply. A priori, by letting the customer explore freely the list of available products, e-commerce should enable him to maximize his satisfaction. Nevertheless, in 60 % of the cases the customer leaves without any purchase and the conversion rate visitor/customer does generally not exceed 15 %. In on-line sale contexts, one of the main limitative factor is the difficulty for the user to focus on products that satisfy his preferences, and in an orthogonal way, the difficulty for the supplier to guide potential customers. This difficulty increases with the size of the e-catalogue, which is typically large when the considered products can be configured: the search space is then combinatorial. The goal of our project is to study how configurators can help a client by guiding his choices, like recommendation systems do, without losing their ability to work on combinatorial domains. This would enable both to perform preference-based guided configurations or collaborative filtering in configuration contexts, and to build recommendation systems which propose the same interactivity than configurators. From a scientific point of view, the originality of this project relies on two main ideas: on the first hand, the use of learning techniques to solve combinatorial problems; on the other hand, the use of compilation approaches, not only for the catalogue or for customer indicators, but also for the recommendation model. This project is a fundamental research project which is planned for 3 years. It involves the Institut de Recherche en Informatique de Toulouse (IRIT), the Laboratoire en Informatique et Robotique et Microélectronique de Montpellier (LIRMM)' and the Centre de Recherches en Informatique de Lens (CRIL)', but is not purely academic. The participation of three industrial partners specialised in configuration softwares,, Cameleon software, IBL and Renault, provide the
project with an expert point of view on actual needs, and a case study. The project deals with two important research domains for e-commerce, namely, the development of online recommender systems and configurators. It aims at developing the advising functionality and taking into account the customer's preferences in configuration-based systems, both with ``classical'' B2B configuration (technical object configuration) or with interactive exploration of a B2C catalogue (preference-based search or interactive exploration). To this end, this project proposes to pool the experience about those two kinds of system, studying jointly learning techniques, collaborative and compilation-based filtering and/or preference propagation. To the best of our knowledge, this approach is absolutely original with respect to the international state of the art techniques. This project is a revised version of the "BR4CP" projet submitted as a "ANR-Blanc" proposal for 2010 - it includes a new partner (IBM) that will (among others) reinforce the analysis and the validation tasks.