Résumé
The works presented in this thesis are mainly based on folksonomies: a collaborative data structure that emerged with Web 2.0. The increased need for collaborative services to index, classify and retrieve information in various forms, the huge amount of data generated as well as the heterogeneity of the data sources has favored the emergence of processes by which users qualify online contents, notably by associating tags (or descriptors) to them. Such processes produce data that can be modeled by hypergraphs considering tripartite graphs representing folksonomies and constituting the set of tags assigned by users to resources. Since these tags are user-generated, they form a new set of personal data that reflects users’ interests and preferences for resources or items on the Web. Tags are used in various information retrieval tasks, specifically in the area of personalized information retrieval and recommendation. For these kinds of tasks, user-specific characteristics are extracted and then processed to enrich folksonomy profiles. These profiles are built using different approaches: vectorial, temporal, or clustering-based. It is in this context that we approached our first contribution, insofar as we were able to implement an unsupervised method for the construction of folksonomy profiles. We found that these profiles, which we represented in the form of evolving interest clusters, were better able to describe users and resources efficiently. They have been successfully implemented in a recommendation process and then evaluated among classical vector profile construction methods. This helped us to better understand the diverse and evolving nature of user interests associated with social annotations. Specifically, by exploiting the normalized degree of user preference propagated by tags. Our works then turned to the study of recent works in representation learning, which exploit neural models to enrich profiling methods with continuous representations of users and items. We found that such models have significantly improved the performance of classical recommendation systems. This led us to consider enriching these neural models by associating user and item representations with lexical embeddings of tags. These tag embeddings that we extracted from pre-trained neural language models allowed us to formalize neural collaborative filtering models. This is the context of our second contribution. After exploiting these representations (embeddings) through different neural architectures, we have highlighted which neural models allow the extraction of more precise characteristics and provide information on the contextual semantics of tags. Finally, the integration of these representations within these neural architectures, also allowed us to address other underlying issues and which are related in particular to the way to effectively include the neighborhood of tag representations in a neural collaborative filtering model for recommendation. In this sense, we have evaluated the accuracy of users’ rating predictions from annotation history and then determined which models lead to better performances compared to classical neural collaborative filtering approaches that do not necessarily integrate this type of representation.
Source: http://www.theses.fr/2022TOU30256
.
