Exploiting Active Learning in the Context of Physio-Adaptive Systems

Problem Description

Problem

Physio-adaptive systems define a class of information systems that refer to an innovative mode in which system interaction is achieved by monitoring, analyzing, and responding to covered human psycho-physiological activity in real time (Fairclough, 2009). Such systems span a range of application areas, such as: health, gaming, aviation, transportation, and learning (Loewe & Nadj, 2020). However, a major challenge in implementing physio-adaptive systems remains the reliable and accurate detection of psycho-physiological states. In agreement with Fairclough (2009), the accuracy of detection of these states is a critical success factor for the adoption and use of such systems. Complicating matters further, physiological data are highly person-specific and individualized, making the need for labeled data for training a classifier even greater compared to other classification problems. While manually creating a large labeled dataset is often a costly, error-prone, and frustrating activity, more efficient approaches such as active learning open up new opportunities to collect labeled data. Active learning represents a subdomain of machine learning where a learning algorithm identifies interesting key data points and interactively queries the user or another source of information for its label. Therefore, only a subset of data has to be labeled. This is specifically beneficial for scenarios where the label collection is time-consuming, expensive, or complex (Nadj et al., 2020).

Goals

The goal of this thesis is to implement an active learning approach for an already existing physio-adaptive system. The investigation, evaluation, and integration of online machine learning is also part of the thesis project.

Requirements

We expect extensive knowledge of procedures within established machine learning methods, such as Random Forest, Support Vector Machines and Neural Networks as well as experience with Active Learning and Online Machine Learning. Profound knowledge of machine learning libraries such as scikitlearn is also required. In addition, experience in pre-processing and analysis of physiological data is beneficial.

Contact

If you are interested in this topic and want to apply for this thesis, please contact Dr. Mario Nadj with a short motivation statement, your CV, and a current transcript of records. Feel free to reach out beforehand if you have any questions.