Modeling users, Understanding context, Adapting recommendations, Improving interaction.

Research Area and Main Goals

Recommender systems are the common focus of the works conducted in the team, since its creation in 2008. KIWI has been the first team in the lab to work on recommender systems. The scientific focus is the automatic exploration of digital traces: logs, clickstreams, ratings, annotations, writing in blogs, etc. This exploration is based on models issued from machine learning, data mining, subjective logic, collaborative and content-based filtering, considering only traces or including human factors for their processing.

Our objective is to model the user behavior (descriptive modeling), explain it (diagnostic analysis), predict its evolution (predictive modeling) or determine what actions to do to achieve a goal (prescriptive modeling). Our research topics include individual (user) and collective modeling (community), instant and dynamic modeling, single domain (cultural goods, educational resources) and cross-domain systems, personalization (adaptation to the user) and flexibility (adaptation to the context). Application domains were mainly e-education, e-health, digital media services, e-commerce, cultural heritage, information systems and social networks.

Human Factors in Information Retrieval

Recommender systems have been proven to be efficient and useful by reducing the cognitive load and time required during data search and access. Over the past two decades, this improvement of human-computer interactions is mainly relying on increasing systems’ accuracy. A crucial aspect is missing within the literature evaluation metrics. They do not take into account human factors playing a role within the decision process (context, confidence, trust, explanations and need for diversity). Within this context, our goal is to design holistic intelligent systems that provide the right information at the right time, in the correct manner, in agreement with users’ policy and with valuable arguments. New challenges consist in: (1) identifying human factors that play a role within decision making an/or maximize users’ acceptance, adoption and satisfaction, (2) integrating these factors in machine learning algorithms.

Predictive Modeling and User Characterizing

This axis is dedicated to user characterization through clustering, representative users (leaders), atypical users, etc. to provide a simplified representation of the set of users, and to provide them with accurate recommendations. It is also dedicated to predicting users future behavior. We mainly focus on machine learning approaches to reach these goals.

Hybrid Modeling

Hybridization aims to combine knowledge sources to design users model to produce personalized recommendations. The key idea is to globally improve the quality of recommendations taking advantage of the specificities of each method. The issues addressed are: cold start problem, robustness to attacks, data sparsity, their fluctuation and massification. Knowledge sources allow to infer relevant information (indicators and features). They can be linked to items including their contents and their uses, but also to users social connections in an online service. This informations can be explicit: ratings, votes, content metadata and declared social links. They may also be implicit: we estimate them from the use of any kind (access, frequency, duration, sharing, downloading, printing ...), the textual content of items or the interactions among users.

Education and Digital

This axis is dedicated to e-education and more precisely to the way to develop personnalized education for any learner, based on the collect and exploration of digital traces. It includes questions such as how to recommend pertinent open educational resources (OERs) to a specific learner given his academic background, educational preferences and learning objectives, or how to adapt digital tools to his profile. The KIWI team maily focuses on designing an OERs recommender and on Learning Analytics (KIWI is involved in this topic since 2013, mainly via the PIA 2 Pericles project) to provide teachers and learners with explicative, predictive and prescriptive analysis.