Personalização de conforto em ambientes inteligentes por transferência de conhecimento em aprendizado profundo

Abstract

Predicting personalized comfort in smart environments can be very beneficial for its users. For example, personalized smart offices can improve the well-being and work performance by optimizing the users' sense of comfort. However, in order to generate comfort prediction models, traditional machine learning algorithms require an amount of labeled data that is expensive in terms of time and financial resources. Smart environments typically generate a low amount of data in an acceptable period of time, and this problem becomes worse with the personalization of comfort. Personalization requires the collection of individual user data, facing problems of user availability and willingness to provide the requested data. Our proposal applies knowledge transfer with fine-tuning to reduce the learning time of a personalized intelligent environment. This technique allows models pre-trained on other tasks to transfer the knowledge obtained to a new task. Our evaluation we performed with three different data sets. The results showed that knowledge transfer improves the performance of a model when compared to training without knowledge transfer. The improvement ranged from 2.24% to 14.82%, with the average improvement being 7.93% across all combinations of target and source users.