PhD Thesis - Multi-Model Semi-Supervised Learning for Personalisation
Using unlabelled data to improve machine-learning model prediction accuracy
Nowadays, the applications and systems generate large amounts of unlabelled data which by itself carry no information. For example, wearable and environmental sensors continuously sense and measure the respective data, speech while talking to the mobile assistants (Siri, Google assistant, Alexa, etc.) is sensed, large amount of new content on websites is generated daily, etc. Labelling such large amounts of data is very labour intensive and expensive. Even though the idea and need for automatic labelling arose decades ago (Scudder 1965, Vapnik and Chervonenkis 1974), this area is still not mature enough to be generally applied.
Semi-supervised learning is a technique that aims at utilising a small set of labelled data to automatically label the unlabelled data. The topic is very popular due to the mentioned fact that in all domains there are cases where large amount of unlabelled data is generated and the need for labelling it exists. There are several approaches to tackle the task but they are either limited to offline labelling or the selection process is implemented with empirical risk minimisation approach, which means that only the most confident predictions are included in retraining and consequently do not contribute effectively to the solution.
The thesis will explore the possibility to employ machine-learning into the selection process and enable inclusion of the predictions which might be discarded using the heuristics approach to empirical risk minimisation. This approach will contribute to inclusion of more diverse data and therefore more effective adaptation of the learner.
PhD Thesis: CVETKOVIĆ, Božidara. Multi-Model Semi-Supervised Learning for Personalisation.PhD Thesis, 2018, International Postgraduate School Jožef Stefan.
CVETKOVIĆ, Božidara, JANKO Vito, ROMERO E. Alfonso, KAFALI Ozfur, STATHIS Kostas, LUŠTREK, Mitja. Activity Recognition for Diabetic Patients Using a Smartphone.Journal of Medical Systems, 2016, vol. 40, no. 12, str. 256-1-256-8.
CVETKOVIĆ, Božidara, KALUŽA, Boštjan, LUŠTREK, Mitja, GAMS, Matjaž. Adapting Activity Recognition to a Person with Multi-Classifier Adaptive Training.Journal of Ambient Intelligence and Smart Environments, vol. 7, no. 2, 2015, pp. 171-185.
The Fit4Work project aims at delivering an innovative system capable of detecting, monitoring and countering physical and psychological stress related to occupation of older adults. The project will extend off-the-shelf technologies (3D motion sensing, wearable wellness sensors, ambient sensors, and mobile devices) with specialized components able to analyse physical and mental fitness in order to provide personalized recommendations and exercises. The resulting product will be a coupling of unobtrusive devices and services that make it possible to continuously monitor self at work and manage own fitness thanks to motivating training scheme. The goal is to increase users’ quality of lives and health-related fitness.
e-Gibalec - Mobile application to monitor and promote exercise in schoolchildren for more effective physical education
The e-Gibalec project is developing a mobile application that will use smartphone sensors and intelligent computer methods to monitor the movement of children. Gamification techniques will then be used to encourage them to be physically more active. The application will also involve parents and physical-education teachers. The teachers will have access to the data collected by the application and will be able to use it for more effective physical education. The application will be tested in multiple primary schools. The long-term goal is to introduce it to all interested primary and secondary schools.
Commodity12 – COntinuous Multi-parametric and Multi-layered analysis Of DIabetes TYpe 1 & 2
COMMODITY12 aims at improving the daily management of diabetes and the prevention/management of its cardiovascular co-morbidities. A multi-layer multi-parametric infrastructure is being developed which monitors the patient's physiological signals and lifestyle, and analyzes patient's data to produce indicators to doctors concerning diabetes and its cardiovascular co-morbidities.
We will contribute to the project by developing methods for monitoring the patient's lifestyle: recognition of patient's elementary activity (e.g. walking, sitting, lying), estimation of patient's energy expenditure and recognition of his/her main daily activity groups (e.g. work, exercise, rest). The reasoning is based on data obtained from an accelerometer placed on the patient's chest and/or from sensors integrated in a smart phone worn by the user.
E-Turist is a mobile application that will attempt to provide a tourist with an experience comparable to that offered by a professional tour guide, but tailored specifically to him/her. The tourist will enter his interests, the available time and any special requirements he/she may have. Based on these, the application prepares a personalized sightseeing program using a recommender system. Afterwards, the application guides the tourist using the GPS, providing a multilingual description accompanied by photos. The description will be available on the mobile phone screen and via synthesized voice. The tourist may comment and rate each sight, which is then used by the recommendar system and tourism workers to improve their services. The mobile application is accmpanied by a web application through which tourism workers may enter information on sights of interest and track the activity of their visitors.
The eTurist application is available for iOS 7, Windows phone, Android and Blackberry.
The aim of the project is to develop a prototype of an electronic doorman, named e-doorman, that offers services similar to a human doorman, improves security and increases user comfort. The intelligent system is embedded into a door with electro-mechanic lock, tablet PC, micro-controller and an array of sensors. The e-doorman uses context-based reasoning and awareness achieved by artificial intelligence methods running on the tablet. The e-doorman system is able to recognize the users, detect unusual entry/exit, break-in attempts, predicts user presence and offers personalized services such as customizable notifications and alarms, information about present residents and state of the door, voice messages, greetings and tips, and remote control using intuitive GUI or virtual assistant that understands natural language.
A free Android application developed as a side product of the project is available.
CHIRON – Cyclic and person-centric Health management: Integrated appRoach for hOme, mobile and clinical eNvironments
The project will develop an integrated framework for personalized healthcare at home, in a nomadic environment and in the hospital. A patient is equipped with wearable sensors, which continuously monitor his condition. The sensors are connected to a smartphone, which issues warnings and advises the patient based on a personalized health assessment model. The data from the sensors, together with the data from the patient's health record and novel medical imaging solutions, is accessible to medical professionals. Their work is supported by an advanced health assessment model, which is continuously modified by incoming data and experts' input. Our department is developing this system, as well as activity recognition and human energy expenditure estimation methods for patient monitoring.
We have developed activity monitoring application, which uses sensor data (accelerometer and heart-rate monitor) to recognise the human activity and estimate the energy expenditure.
CONFIDENCE - Ubiquitous Care System to Support Independent Living
The main objective of project Confidence (Ubiquitous Care System to Support Independent Living) is the development and integration of innovative technologies to build a care system for the detection of abnormal events (such as falls) or unexpected behaviours that may be related to a health problem in elderly people.