REPOSITORY > RESULTS

Abstract

In this dissertation, we propose a novel semi-supervised learning method called Multi-
Model Adaptive Training for adaptation and personalisation of machine-learning models.
The method is designed to efficiently exploit large amounts of unlabelled data generated
by wearable devices while they are in use.
Systems that utilise machine learning are usually deployed with a single or multiple
machine-learning models trained with datasets collected in a controlled environment and
explicitly labelled for their task. Once such models are used in real life, they may perform
well if the statistical properties of the data are stable and similar to the training
dataset. If the statistical properties of the data vary or drift, the models might diminish
in accuracy since the collected dataset on which they were trained may not include all
possible variations. The solution is either to collect and hand-label data containing new
statistical properties and train the model in a supervised manner, which is time-consuming
and expensive, or to use a methodology that can exploit unlabelled data. The latter approach
is cheaper since an overwhelming amount of unlabelled data is generated while
such systems are in use. To enable this data to be used for personalisation it needs to be
labelled either by a human annotator when required, via active learning, or automatically,
via semi-supervised learning (SSL). The latter is the subject of this dissertation, in which
we propose a method that builds on the idea of probabilistic methods called Self-training
from the SSL low-density separation methods group.
The idea of Self-training methods is to train a model or an ensemble of models in a
supervised manner and use them to label unlabelled examples. The decision whether or
not an automatically labelled example is included in the training set used for retraining
is done with heuristics, which are typically designed to include only the most confident
predictions to avoid a potential decrease in accuracy. However, using only highly confident
predictions in retraining may not improve the performance compared to the initial model,
since the included examples already fit it well.
The proposed method increases the potential for improvement after retraining by also
including examples labelled with a higher uncertainty, which is achieved by replacing the
heuristics with a machine-learning model. Each unlabelled example is first labelled with
two domain models, a general model trained on data of other people than the target person
and a person-specific model trained on a small amount of data specific for the target person.
The results of the domain models are stacked with a selection model that selects which
domain model should provide the final labels for the example. The results of the domain
models and the selection model are additionally processed with an inclusion model trained
to decide whether the automatically labelled examples should be included in the training
set for retraining.
We evaluate the personalisation using the proposed method on three classification tasks
and one regression task. We analyse each step of the method and compare it against other
SSL approaches. We achieve successful personalisation and show that the proposed method
outperforms competing approaches by up to 10 percentage points on average.