Background - Smartphone-Based Human Activity Recognition - page 20

adaptive and flexible in accordance to the new inputs. This type of learning is required
in applications with high output variability and where a stream of new samples is
available and can be progressively added to the model for learning. This is the case
of online web ranking and stock market prediction applications.
2.5.1.2 Unsupervised Learning
In an unsupervised learning problem, the training data consists of only input vec-
tors without their associated targets . It aims to find certain similarities or discover
distinguishable structure within the input data (e.g. clustering ). It can also be used
for density estimation to describe the distribution of the data in its space. Moreover,
this learning approach can be exploited for data visualization using dimensionality
reduction methods which allow to better project high-dimensional data into smaller
spaces (e.g. 2D and 3D).
Unsupervised learning approaches have been already applied in several areas such
as in medical imaging where 3D Positron Emission Tomography (PET) scans use
cluster analysis to find dissimilarities between different organs and types of tissue to
be able to correctly segment the scanned area (George et al. 2011 ). It has also been
applied in the automatic grouping of similar shopping items (e.g. books, movies,
music), particularly in recommender systems for online stores that aim to predict the
user preferences based on products similarities and previous purchases.
2.5.1.3 Semi-supervised Learning
This learning approach combines labeled and unlabeled data for learning. Therefore,
it takes aspects from both supervised and unsupervised approaches. In general, small
amounts of labeled data are integrated with a large number of unlabeled samples for
learning. For example, it is useful for datasets where it is not always possible to have
a label for each sample. Evidence have shown that semi-supervised learning can
greatly improve the learning performance when compared with supervised learning
which does not take into account unlabeled data. This is feasible if considerations
such as the data smoothness assumption apply (Chapelle et al. 2006 ). Vast digital
image collections on the internet for content retrieval are an application example
where this type of learning can be exploited. Not all the images have an associated
targets and it would be humanly impossible to perform this labelingmanually (Fergus
et al. 2009 ).
2.5.1.4 Reinforcement Learning
This learning approach is oriented on finding an appropriate set of actions to solve a
particular problem. This is done with the purpose of maximizing a reward. Optimal
solutions are not found through learning a model given a set of input-target pairs.