Résumé
User data is becoming increasingly available in multiple domains ranging from e-commerce platforms to social media networks. It includes demographics (e.g., age, gender, location, etc.) and user activities (e.g., browsing habits, purchase history, rating records, etc.). The analysis of this data is appealing as it helps companies to enhance their business, understand users’ behavior, reduce their churn, and attract new customers. With the advancement of technology, many data-driven and data-informed tools were developed to understand the preferences of users and extract valuable insights from the collected data.
In this thesis, we propose a framework that aims to study distinctly both the individual and collective behavior of users. We first aim to examine users individually as each one is unique, and their actions and interactions may vary significantly from one to another. We leverage recommender systems that analyze the behavior of users at a low level of granularity which allows for personalized experiences that can better meet users’ expectations. More precisely, we leverage dynamic recommenders to infer the states where the users might be and capture their constant evolution over time. Following this, we first extend the standard recommenders by incorporating users’ states and profiles within a static environment based on a meta-learning methodology. Then, we explore more realistic contexts where the environment is dynamic. We explore three real-world applications: Educational test recommendation, SQL query recommendation, and diverse session recommendation. Within each application, we define the behavior of users with many dimensions to avoid overspecialization and filter-bubble and propose several solutions based on Multi-armed bandits, and Reinforcement Learning.
In addition to their unique behavior, users with the same characteristics (e.g., demographics) may exhibit the same global trends. Hence, our second part of the framework aims to extract these insights, seek to analyze users’ collective behavior, and discover the relationships between the different user groups and the subset of items. For the purpose of reducing false discoveries, we rely on hypothesis testing to produce significant and statistically sound insights. We also optimize for coverage to explore all users’ groups and avoid analyzing a small subset of them. We design novel solutions based on standard multiples hypothesis testing corrections as well as α-investing.
In this thesis, we evaluate our solutions using an extensive set of experiments both for quality and performance. We also conduct a comparative analysis with existing approaches or state-of-the-art approaches to demonstrate the effectiveness of our solutions.
Source: http://www.theses.fr/s256977
.
