measuring the similarity between two MTS items is also important
for classifi cation. In practice, Euclidean distance [20] and dynamic
time warping (DT W) [21] are two of the most popular methods.
The former is a fast method, but its quality is easily affected by
abnormal data points, while it often requires that the time series
are equal in length. DTW is a robust method, but the time and the
space complexity of its computation are very large, which make it
unsuitable for long time series with excessive numbers of
variables.
In most cases, MTS hav e differ ent lengths. Most of the traditional
techniq ues can reduce the variable-based dimensions but the number
of time-based dimensions is retained; thus, MTS with different
lengths will yield sequences that represent the MTS with different
lengths. For instance, PCA transforms MTS into the corresponding PCS
with different lengths, which means that distance functions such as
DTW must be used to measure the similarity . However, although
DTWisaneffectiveapproach,itrequiresexcessiveamountsoftime
and space to measure distances.
To overcome the problems mentioned above, we propose an
accurate and efficient MTS classification method. The main moti-
vations of our study are summarized as follows. First, we analyze
the weaknesses of the traditional common PCA (CPCA) applied to
the field of time series data mining. The effectiveness and the
efficiency of CPCA are regarded as shortcomings of the algorithms
used to extract knowledge from MTS datasets; thus, it is necessary
to design a novel method to address these difficulties. Second, the
traditional methods based on PCA often fail to handle MTS data
with different lengths; therefore, the proposed method should
consider various lengths and improve the quality of PCA for
mining time series.
In this study, various MTS clusters are transformed to construct
the corresponding reduced subspaces by CPCA. Each space is then
organized by the eigenvectors of the common covariance matrix in
a cluster. The MTS without class labels in the test dataset are
projected onto the different subspaces and the minimal variance in
the reduced PCS according to different projections can specify the
label values for the MTS in the test dataset. The two main
contributions of our proposed method are as follows. First, MTS
items with the same label in a cluster are used to construct the
subspace, which means that the PCS of any MTS item projected
onto the corresponding subspace has a large variance. Thus, when
the variance in the PCS derived from different subspaces is larger,
the projected item and those in the corresponding cluster are
more similar. Second, we treat the largest variance in the different
subspaces as a classifier with high efficiency, which improves the
performance of the proposed method. Three advantages may be
obtained, as follows. (1) Our proposed method is faster at
classifying MTS compared with the existing methods based on
PCA. (2) The quality of the classification results obtained by the
proposed method is often better than that with traditional
method. (3) The proposed method is suitable for the classification
of MTS datasets where the lengths of the MTS items are different.
The results of our experimental evaluation also indicated that the
proposed method is more accurate and efficient.
The remainder of this paper is organized as follows. Back-
ground and related work are introduced in Section 2.InSection 2,
we describe the proposed new classification method. The results of
experimental evaluations of the proposed method are presented in
Section 4. In the final section, we give our conclusions and discuss
future work.
2. Background and related work
Due to the high dimensionality of MTS, techniques for dimen-
sionality reduction are very important for time series data mining,
and PCA [15,17,10] is one of the most commonly used methods. In
addition, compared with the traditional methods, CPCA can often
improve the performance of the algorithms used in MTS datasets.
In this section, we introduce both these methods and we review
related work.
2.1. PCA
PCA is used widely for MTS dimensionality reduction. PCA can
transform a MTS X ¼fx
1
; x
2
; …; x
m
g into a PCS Y ¼fy
1
; y
2
; …; y
m
g,
where m is the number of variable-based dimensions and y
i
denotes the ith principal component sequence. Moreover, the first
principal component sequence y
1
contains most of the informa-
tion about the original MTS and y
2
contains the second highest
amount of information, and so on. In fact, each principal compo-
nent sequence y
i
is a linear transformation of the variables in the
original MTS and the coefficients defined in this transformation
are considered as weight vectors, i.e.,
y
i
¼ a
1i
x
1
þa
2i
x
2
þ⋯ þa
mi
x
m
; i ¼ 1; 2; …; m; ð1Þ
where a is the corresponding weight and x
j
denotes the jth
variable of the MTS. Moreover, the first principal component
sequence y
1
has the largest variance, λ
1
¼ Varðy
1
Þ, the second
principal component sequence accounts for the largest portion of
the remaining variance, λ
2
¼ Varðy
2
Þ, and so on. In this manner, the
first p component sequences may retain most of the variance
present in all of the original m variables, where po m. Thus, the
dimensionality reduction for a MTS with m variables can be
achieved by projecting it onto the p- dimensional subspace (also
called the coordinate space). The subspace can be constructed by
an eigenmatrix of the covariance matrix Σ of X. According to the
SVD, the covariance matrix can be decomposed by
Σ ¼ UΛU
T
; ð2Þ
where U contains the weights for the principal component
sequences and the matrix Λ has the corresponding variances,
which means that the first column vector of U is the weight vector
of the first principal component sequence and its variance is the
first element of the matrix Λ along the diagonal.
To reduce the dimensionality of MTS, the first p principal
component sequences are retained, which means that the first p
eigenvectors are used to construct the subspace, i.e.,
A
mp
¼ Uð1 : m; 1 : pÞ. Thus, the reduced PCS can be formed by
Y
np
¼ X
nm
A
mp
; ð3Þ
where n is the length of the MTS, where it is often the case that
po m and po n. In this manner, we can transform a MTS with a
size of n m into a reduced representation with a size of n p.
In addition to dimensionality reduction and feature represen-
tation using PCA, some distance functions are often used to
measure the similarity between two representations following
the transformation of the MTS by PCA. The angles between all the
combinations of the selected principal components can be used to
measure the similarity [22]. Another approach was proposed by
[23] for modifying previous methods by weighting the angles with
the corresponding variances. Ref. [17] addressed the issue of
similar principal components present in a time series by using
the different values of the variables. Refs. [24,10] proposed Eros
based on the acute angles between the corresponding compo-
nents, which can measure the similarity better and faster than
previous methods. A fast similarity search for MTS using a
projection comparison based on PCA was proposed by Karamito-
poulos et al. [25]. In addition, since PCA is based on SVD, some
methods based on SVD [11,12] have been applied to dimension-
ality reduction and as similarity measures for MTS.
H. Li / Neurocomputing 171 (2016) 744–753 745