Abstract
This thesis presents a study that aims to develop a unified framework for the design of production systems for the mass personalisation of skin-care. It addresses the application of machine learning and more specifically supervised learning techniques in customisation of cosmetic formulations through personalising their ingredients. The proposed framework takes into consideration the limitations associated with data gathering through medical trials and proposes solutions to maximise knowledge acquisition from the valuable biomedical data. It also rectifies the shortfalls of the expert systems that are often used for personalisation of medical and dermatological products. The framework incorporates a main production unit that employs a supervised learning agent to gain knowledge through learning the connection between human skin-profile and ingredients of personalised skin-care products. After its training stage, the unit is able to produce personalised skin-care with an accuracy proportionate to its initial knowledge source.The thesis also introduces an adaptation unit to further improve on the generated
skin-care formulations by incorporating the product feedback provided by the
patients. To do so, a novel iterative feedback learning technique is introduced. This technique utilises an independent supervised learning agent to recognise the relation between the patients’ skin-profile, formulations and provided feedback. It adjusts the levels of ingredients in the produced formulations to generate more effective versions and to acquire more favourable feedback from the patients.
The framework introduces the concept of knowledge transfer from the adaptation unit to the main production unit to improve its knowledge of personalised skin-care products through incremental learning. The system built under this framework requires initial supervision and training. However, it is able to self-update by utilising patients’ feedback after the initial training stage. The thesis introduces multiple task groups to evaluate modules of the proposed framework and the introduced novel techniques. The task groups show constant improvement in generated formulations by utilising the proposed adaptation module. Up to 32% improvement has been reported in the accuracy of the formulations only after receiving three feedbacks from the patients for their favourite skin-care products. Additionally, the impact of different types of skin-measurements on the generated formulations is investigated in this work.
| Date of Award | 15 Jul 2019 |
|---|---|
| Original language | English |
| Supervisor | Branislav Vuksanovic (Supervisor) |