Simulation eines Reinforcement Learning Model zur Anwendungen in Just-In-Time Adaptive Intervention Recommender Systems

Katharina Brunner

Publikation: AbschlussarbeitMasterarbeit

Abstract

Globally speaking, consistent healthcare and easy access to health services for all citizensis still a utopian concept. In some areas of this world the coverage of mobile networkssurpasses the local health care infrastructure. In order to combat this issue, digital healthinitiatives have emerged in recent years, promising possible solutions. An important butcomplex sector in digital health is the development of treatment recommender systems,which are machine learning driven multi-component applications that utilise artificial intelligenceto deliver personalised supportive intervention to a client, based on just-in-timeintervention delivery, and adaptiveness.This thesis looks at a machine learning algorithm called Thompson Sampling with RestrictedContext, or TSRC, and investigates whether it is a contender for the engine in ajust-in-time adaptive recommender system. First, an overview of the framework for medicalrecommender systems is given, which includes a description of its key elements and adiscussion of the current applications working with this concept.Mathematically, the problem faced by recommender systems can be interpreted as acontextual multi-armed bandit problem. The standard algorithms solving this problem arepresented, and their advantages and disadvantages are discussed before arguing why theThompson Sampling approach is selected for this thesis.Subsequently, Thompson Sampling is investigated as a machine learning paradigm, andthe TSRC algorithm is presented as an extension of the traditional heuristic, which, due toits restricted context policy may be equipped to handle cases where contextual informationis missing, for example in the case of a technical failure to record data.In order to analyse the TSRC algorithm’s performance in choosing supportive interventions,a reinforcement learning-based model is designed and implemented in Matlab. Itincludes a model recommender system together with virtual model clients, and the implementationof the TSRC algorithm.Thereafter, simulations are performed with the model recommender system and differentclients, and the TSRC algorithm’s response to contextual feature sparsity and cases ofmissing data, both relating to restricted context, are investigated. All simulation resultsstrongly suggest that the TSRC algorithm is a contender for just-in-time adaptive recommendersystems, and an outlook containing future research into the topic is provided. enter abstract of the article in the language of the published article].
OriginalspracheEnglisch
Gradverleihende Hochschule
  • TU Wien
Betreuer/-in / Berater/-in
  • Hametner, Bernhard, Betreuer:in
  • Breitenecker, Felix, Betreuer:in, Externe Person
Datum der Bewilligung17 Jan. 2022
DOIs
PublikationsstatusVeröffentlicht - 2022

Research Field

  • Medical Signal Analysis

Fingerprint

Untersuchen Sie die Forschungsthemen von „Simulation eines Reinforcement Learning Model zur Anwendungen in Just-In-Time Adaptive Intervention Recommender Systems“. Zusammen bilden sie einen einzigartigen Fingerprint.

Diese Publikation zitieren