18
Internet of Things (IoT)
2.1 Introduction
Supervised and unsupervised learning techniques are reliable tools for the classification
and categorization of data. Wearable devices are a relatively recent consumer technology
that record large amounts of data by measuring signals around the body. Wearables come
in many different forms, such as wristbands, armbands, watches, and headbands. This
chapter explores the use of supervised and unsupervised learning techniques to identify
individuals and activities using a commercially available wearable headband.
Wearable headbands typically measure the electroencephelogram (EEG) signals generated by the user’s brain from many different locations around the head. Depending on the
duration of measurement, these signals generate a large stream of values of variable length
on multiple channels, from multiple locations. It is not easy to use the raw representation
generated by such devices for meaningful data mining activities. This chapter illustrates
a method of creating a compact representation of the data streams from multiple channels
without losing the essence of the patterns within the data.
The data were collected from a commercially available wearable headband, called Muse.
The Muse headband records EEG signals from four different locations around the head,
as well as acceleration and some facial movement data. The dataset was created based on
the recordings of five individuals performing five tasks each; reading, playing computer
games, relaxing, listening to music, and watching movies. The activities were repeated
a number of times for each participant. The raw data were converted to the proposed
knowledge representation for use with the various supervised and unsupervised learning
techniques.
The viability of the proposed knowledge representation is demonstrated through the
usage of a number of well-known supervised classification [1] techniques, including decision
trees [2], support vector machines (SVM) [3], neural networks [4], and random forests [5].
The usage of a variety of classification techniques show the summarized frequency distribution is effective for representing the time series of signals, independent of the classification
techniques that are used. The classification techniques successfully predicted the persons as
well as the activities based on the data.
This chapter also explores the usage of unsupervised and semi-supervised learning
techniques with the proposed data representation. Clustering using the K-means algorithm is one of the most popular unsupervised learning techniques. K-medoids is an
alternative of K-means that finds an object that is most similar to all the other objects in
the cluster, as opposed to determining the centroid of the cluster. These methods focus on
optimizing within cluster scatter and separation between the clusters. Both K-means and
K-medoids do not provide the capability of incorporating additional optimization criteria. The fact that K-medoids offers a discrete search space, limited by the total number
of objects in the dataset, can provide advantages for evolutionary searching. By combining K-medoids with an evolutionary algorithm, it is possible to perform multi-objective
clustering. Peters [6] first proposed the use of evolutionary computing in the context
of rough set theory, Lingras [7,8] subsequently explored both K-means and K-medoids
based on rough set theory.
The K-means and the proposed semi-supervised crisp and rough K-medoid algorithms are compared using the proposed knowledge representation structure. By extending the evolutionary rough K-medoid algorithm to optimize the precision of the known
categorization of signals in the dataset, it is shown that this approach may be effective for
improving the precision of known category information in some cases.
Internet of Things (IoT)
2.1 Introduction
Supervised and unsupervised learning techniques are reliable tools for the classification
and categorization of data. Wearable devices are a relatively recent consumer technology
that record large amounts of data by measuring signals around the body. Wearables come
in many different forms, such as wristbands, armbands, watches, and headbands. This
chapter explores the use of supervised and unsupervised learning techniques to identify
individuals and activities using a commercially available wearable headband.
Wearable headbands typically measure the electroencephelogram (EEG) signals generated by the user’s brain from many different locations around the head. Depending on the
duration of measurement, these signals generate a large stream of values of variable length
on multiple channels, from multiple locations. It is not easy to use the raw representation
generated by such devices for meaningful data mining activities. This chapter illustrates
a method of creating a compact representation of the data streams from multiple channels
without losing the essence of the patterns within the data.
The data were collected from a commercially available wearable headband, called Muse.
The Muse headband records EEG signals from four different locations around the head,
as well as acceleration and some facial movement data. The dataset was created based on
the recordings of five individuals performing five tasks each; reading, playing computer
games, relaxing, listening to music, and watching movies. The activities were repeated
a number of times for each participant. The raw data were converted to the proposed
knowledge representation for use with the various supervised and unsupervised learning
techniques.
The viability of the proposed knowledge representation is demonstrated through the
usage of a number of well-known supervised classification [1] techniques, including decision
trees [2], support vector machines (SVM) [3], neural networks [4], and random forests [5].
The usage of a variety of classification techniques show the summarized frequency distribution is effective for representing the time series of signals, independent of the classification
techniques that are used. The classification techniques successfully predicted the persons as
well as the activities based on the data.
This chapter also explores the usage of unsupervised and semi-supervised learning
techniques with the proposed data representation. Clustering using the K-means algorithm is one of the most popular unsupervised learning techniques. K-medoids is an
alternative of K-means that finds an object that is most similar to all the other objects in
the cluster, as opposed to determining the centroid of the cluster. These methods focus on
optimizing within cluster scatter and separation between the clusters. Both K-means and
K-medoids do not provide the capability of incorporating additional optimization criteria. The fact that K-medoids offers a discrete search space, limited by the total number
of objects in the dataset, can provide advantages for evolutionary searching. By combining K-medoids with an evolutionary algorithm, it is possible to perform multi-objective
clustering. Peters [6] first proposed the use of evolutionary computing in the context
of rough set theory, Lingras [7,8] subsequently explored both K-means and K-medoids
based on rough set theory.
The K-means and the proposed semi-supervised crisp and rough K-medoid algorithms are compared using the proposed knowledge representation structure. By extending the evolutionary rough K-medoid algorithm to optimize the precision of the known
categorization of signals in the dataset, it is shown that this approach may be effective for
improving the precision of known category information in some cases.