Filtrage collaboratif en ligne: application à la publicité programmatique

par Sumit Sidana

Projet de thèse en Mathématiques et Informatique

Sous la direction de Massih-Reza Amini.

Thèses en préparation à Grenoble Alpes , dans le cadre de Mathématiques, Sciences et technologies de l'information, Informatique , en partenariat avec LIG - Laboratoire d'Informatique de Grenoble (laboratoire) et de AMA (equipe de recherche) depuis le 01-12-2015 .


  • Résumé

    Internet advertising has become a major economic challenge for online selling companies that have to optimise their catalogs in real time, in order to propose to users products that fit the best to their interests and preferences. Conventional collaborative filtering engines cannot be applied in this context as there are very few ordinal information, corresponding mainly to past purchases, linking users to the products of their interests. In the other hand relevant information, mainly quantitative, can be obtained from the analyses of user sessions. The particularity of past purchases and the real-value indicators obtained from such analyses is their interdependency and also their nature, as ordinal outputs can be considered as ground truth while real-valued outputs are predicted from user profiling tools and may contain noise. The main objective of this project is to develop the new generation of recommendation system for online advertisement that can handle both ordinal and quantitative information, associating users to products. To attain this goal, we first propose to establish a theoretical framework for multi-target learning for recommendation systems that takes into account the interdependencies between the two types of outputs, relating users to products, and also their nature. The interdependency between the outputs could be studied using the hierarchical decomposition of products and a statistical tool for empirical processes extending the Hoeffding inequality. From this analysis we expect to obtain a transformation allowing to reduce the primal multi-target learning problem into simpler dual multi-label classification and regression problems that will path the way for the development of new generation recommendation engines. The user profiling has to be done using the very few labeled data available and by exploiting the graph of users having the same interest, this requires the development of a graph-based semi-supervised model to propagate information over similar users. The overall model needs to be scalable having very low recommendation time. The thesis is a common project with Kelkoo and Best of Media which will provide the data necessary for developing the framework and the models. For this position, we are looking for highly motivated people, with a passion to work in machine learning, information retrieval and the skills to develop algorithms for prediction in real-life applications. We are looking for an inquisitive mind with the curiosity to use a new and challenging technology that requires a rethinking visual processing to achieve a high payoff in terms of speed and efficiency.

  • Titre traduit

    Dynamic collaborative filtering for on-line advertising


  • Résumé

    Internet advertising has become a major economic challenge for online selling companies that have to optimise their catalogs in real time, in order to propose to users products that fit the best to their interests and preferences. Conventional collaborative filtering engines cannot be applied in this context as there are very few ordinal information, corresponding mainly to past purchases, linking users to the products of their interests. In the other hand relevant information, mainly quantitative, can be obtained from the analyses of user sessions. The particularity of past purchases and the real-value indicators obtained from such analyses is their interdependency and also their nature, as ordinal outputs can be considered as ground truth while real-valued outputs are predicted from user profiling tools and may contain noise. The main objective of this project is to develop the new generation of recommendation system for online advertisement that can handle both ordinal and quantitative information, associating users to products. To attain this goal, we first propose to establish a theoretical framework for multi-target learning for recommendation systems that takes into account the interdependencies between the two types of outputs, relating users to products, and also their nature. The interdependency between the outputs could be studied using the hierarchical decomposition of products and a statistical tool for empirical processes extending the Hoeffding inequality. From this analysis we expect to obtain a transformation allowing to reduce the primal multi-target learning problem into simpler dual multi-label classification and regression problems that will path the way for the development of new generation recommendation engines. The user profiling has to be done using the very few labeled data available and by exploiting the graph of users having the same interest, this requires the development of a graph-based semi-supervised model to propagate information over similar users. The overall model needs to be scalable having very low recommendation time. The thesis is a common project with Kelkoo and Best of Media which will provide the data necessary for developing the framework and the models. For this position, we are looking for highly motivated people, with a passion to work in machine learning, information retrieval and the skills to develop algorithms for prediction in real-life applications. We are looking for an inquisitive mind with the curiosity to use a new and challenging technology that requires a rethinking visual processing to achieve a high payoff in terms of speed and efficiency.