Thesis of Felix Bölz

Explainable and Privacy-Preserving Habit-Change Recommendation: Application to Obesity Care


Amazon, BlaBlaCar, Netflix, Booking, Facebook…: all these applications and e-commerce sites make a very intensive use of recommender systems. Indeed, recommender systems actually lie at the core of platform economy and are key enablers for all user-centric applications. They have been intensively studied over the last two decades and have reached a remarkable effectiveness. Thus, recommendations have such an importance in a platform like Netflix, that one should consider that “Everything you see on Netflix is a recommendation”.

Though extremely effective, most recommender systems yet suffer from a number of drawbacks and limitations which recently raised the scientific community interest: diversity and fairness (recommender systems tend to promote a small part of the contents and to maintain the user in a “filter bubble”), transparency and privacy (recommendation decisions are usually opaque and some recommenders collect very sensitive user’s data in a way that the user is not sufficiently aware of), explainability and interpretability (the user should be able  to “understand” why a specific recommendation is proposed to her). If each of these properties has been independently addressed, no holistic approach, able to enforce all, or at least a large subset, of these properties has been proposed so far.

However, their joint effective enforcement is becoming a mandatory feature, especially when dealing with sensitive data and/or when recommendations are used in a critical domain. In such contexts, privacy-preservation, transparency, explainability and interpretability are mandatory elements for making the recommender system trustworthy for, and therefore acceptable by, the user.

Habit-change recommendation in the context of obesity care is such a sensitive and critical domain. Obesity is one of the most prevalent chronic diseases in western countries, where it concerns 1 in 5 adults (source OECD) and even almost 1 in 4 in Germany (23,6%). Its impact in terms of public health is extremely important. Sustainable obesity care and obesity prevention need, as a key action, to make patients change their nutritional and life (e.g., physical activity) habits. In this perspective, recommender systems can definitely play an important role as they allow collecting patient data, aggregating and processing this data, computing a patient’s dynamic profile, and making effective recommendations regarding eating and physical activity.

In this context, this thesis aims at proposing a personalized habit recommender system for obese patients while jointly enforcing the interpretability of the recommendations and the patient’s data privacy. Interpretability and privacy are indeed mandatory properties as: (i) the patient’s health is at stake and recommendations are part of the patient’s care plan, so one should be able to explain and assess proposed recommendations; (ii) the patient’s data that are collected can be very sensitive (e.g., health parameters, mobility traces, social activities…) and must be protected from unintended accesses, even from the medical team; (iv) finally, interpretability and privacy are key properties for guaranteeing the transparency of the recommendations and, in fine, arouse trust in and acceptability of the recommender system.

This means that recommendations should be: (1) relevant for the patient’s contextual situation at the time of the recommendation; (2) enriched with contextualized and personalized explanations; and, (3) implemented in a privacy-preserving way, so that raw patient’s data are protected from unintended access, and both recommendations and explanations do not allow one to infer protected sensitive patient’s data.

Enforcing these properties will require to investigate the a priori contradiction, and the trade-off to make, between recommendation accuracy and explainability on one side, and privacy on the other side. If the state of the art proposes works on accuracy/privacy trade-off, no study has up to now integrated the explainability dimension. However, this trio of properties are mandatory to sustain the trust in the recommender system. As so, beyond e-health, they will definitely shape future personalized recommender systems, whatever the application field.

The thesis will be held in close collaboration with the Adipositas Zentrum, a specialized hospital of Passau, Germany.



Advisor: Lionel Brunie
Coadvisor: Diana Nurbakova
Codirection: Sylvie Calabretto
Cotutelle: Armin Gerl, Harald Kosch