Skip to Main content Skip to Navigation

Industrial Recommendation Systems in Big Data Context: Efficient Solutions for Cold- Start issues

Abstract : Nowadays, users are inundated with the large volume of data flowing in the web. This information overload makes users unable to locate and to find the wanted information at the right moment, especially when they lack sufficient experience to deal with these immense amounts of data. Lately, sophisticated tools were developed to deal with these emergent challenges, among them we find Recommender Systems ( RSs). This thesis deals with automatic RSs that aim to provide items that fit with users’ preferences. These tools are increasingly used by many content platforms to assist users to access to the needed information. In fact, to perform correctly a RS needs to model user’s profile by acquiring information about user’s interests as the visited content and/or user’s actions (clicks, comments, etc). However, modeling these interests is considered as a hard task, especially when the RS has no prior knowledge about a user or an item (cold-start issue). Therefore, modeling user’s profile is complex, since the generated recommendations are often far away from the real user’s interests. In addition, existing approaches are unable to ensure good performance on platforms with high traffic and which host a huge volume of data. In order to solve these issues and to obtain more relevant recommendations, in this thesis we made three main contributions: 1) proposing a CCSDW method to compute criteria and items weights to be used during the profiling of new users to better understand their preferences and to tackle the new-user issue 2) presenting a hybrid RS based on a linear combination of CF and an enhanced CB approach using HFSM to compensate the lack of data about new items and to deal also with long tail issue 3) implementing a distributed recommendation engine with Apache Spark to enhance the scalability and response time. To demonstrate the interest of the proposed approaches, they were evaluated using different data sets in term of coverage and recommendation accuracy. Furthermore, the distributed algorithms were evaluated to validate their scalability in an industrial context.
Complete list of metadatas
Contributor : Ferdaous Hdioud <>
Submitted on : Thursday, April 19, 2018 - 1:12:33 AM
Last modification on : Thursday, April 19, 2018 - 5:44:40 PM
Long-term archiving on: : Tuesday, September 18, 2018 - 12:18:39 PM


rapport_thèse_corrigé-BONNE ...
Files produced by the author(s)


  • HAL Id : tel-01770414, version 1


Ferdaous Hioud. Industrial Recommendation Systems in Big Data Context: Efficient Solutions for Cold- Start issues. Artificial Intelligence [cs.AI]. Faculté de Sciences et Techniques de Fès (FSTF), 2017. English. ⟨tel-01770414⟩



Record views


Files downloads