Evaluating Quantum Support Vector Regression Methods
for Price Forecasting Applications
Horst Stühler
1 a
, Daniel Pranji
´
c
2 b
and Christian Tutschku
2 c
1
Zeppelin GmbH, Graf-Zeppelin-Platz 1, 85766 Garching, Germany
2
Fraunhofer IAO, Nobelstraße 12, 70569 Stuttgart, Germany
Keywords:
Price Forecasting, Machine Learning, ML, Quantum Machine Learning, QML, SVR, QSVR.
Abstract:
Support vector machines are powerful and frequently used machine learning methods for classification and
regression tasks, which rely on the construction of kernel matrices. While crucial for the performance of this
machine learning approach, choosing the most suitable kernel is highly problem-dependent. The emergence
of quantum computers and quantum machine learning techniques provides new possibilities for generating
powerful quantum kernels. Within this work, we solve a real-world price forecasting problem using fidelity and
projected quantum kernels, which are promising candidates for the utility of near-term quantum computing. In
our analysis, we examine and validate the most auspicious quantum kernels from literature and compare their
performance with an optimized classical kernel. Unlike previous work on quantum support vector machines,
our dataset includes categorical features that need to be encoded as numerical features, which we realize by
using the one-hot-encoding scheme. One-hot-encoding, however, increases the dimensionality of the dataset
significantly, which collides with the current limitations of noisy intermediate scale quantum computers. To
overcome these limitations, we use autoencoders to learn a low-dimensional representation of the feature space
that still maintains the most important information of the original data. To examine the impact of autoencoding,
we compare the results of the encoded date with the results of the original, unencoded dataset. We could
demonstrate that quantum kernels are comparable to or even better than the classical support vector machine
kernels regarding the mean absolute percentage error scores for both encoded and unencoded datasets.
1 INTRODUCTION
Machine learning (ML) methods of any kind are
used to generate satisfying and reliable price fore-
casting applications for different use cases and indus-
tries. One example is the sector of heavy construc-
tion equipment dealers that rely heavily on accurate
price predictions. Determining their fleet’s current
and future residual value allows construction equip-
ment dealers to identify the optimal time to resell
individual pieces of machinery (Lucko et al., 2007;
Chiteri, 2018). Using ML methods to calculate the
residual value of construction equipment is of high
interest and has already been tested in the past (Zong,
2017; Chiteri, 2018; Miloševi
´
c et al., 2021; Shehadeh
et al., 2021; Alshboul et al., 2021; Stühler. et al.,
2023). It has been shown, that the use of existing ML
a
https://orcid.org/0000-0002-7638-1861
b
https://orcid.org/0009-0007-5307-7377
c
https://orcid.org/0000-0003-0401-5333
and automated machine learning (AutoML) methods
generate good results for different applications and
datasets (Zöller et al., 2021; Zoph and Le, 2016; Stüh-
ler. et al., 2023).
Although generating desirable results, these meth-
ods have a strong demand for computational power
needed to create accurate results within a reasonable
runtime. Here is where quantum computers (QCs)
with their promised quantum advantage (Shor, 1999;
Harrow et al., 2009; Huang et al., 2021; Liu et al.,
2021) come into play. The new field of quantum ma-
chine learning (QML) based on QCs provides new
possibilities for generating powerful quantum-based
ML applications with better accuracy and less time
and power consumption.
A commonly employed approach in ML for both
classification and regression tasks is the Support
vector machine (SVM) (Steinwart and Christmann,
2008). SVMs depend on the creation of kernel matri-
ces. Selecting the most appropriate kernel depends on
the specific problem at hand, and QML methods pro-
376
Stühler, H., Pranji
´
c, D. and Tutschku, C.
Evaluating Quantum Support Vector Regression Methods for Price Forecasting Applications.
DOI: 10.5220/0012351400003636
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 16th International Conference on Agents and Artificial Intelligence (ICAART 2024) - Volume 3, pages 376-384
ISBN: 978-989-758-680-4; ISSN: 2184-433X
Proceedings Copyright © 2024 by SCITEPRESS Science and Technology Publications, Lda.
vide new possibilities for generating potent quantum
kernels (Huang et al., 2021; Thanasilp et al., 2022).
This study uses various fidelity and projected
quantum kernel techniques to address an industrial
price forecasting application (Stühler. et al., 2023).
Our analysis focuses on evaluating the most promis-
ing quantum kernels documented in the literature and
assesses their performance against an optimized clas-
sical SVM kernel. Similar to the work presented in
(Grossi et al., 2022), we use different feature set com-
binations of the dataset to examine the importance of
the features and their impact on the results.
Our work extends previous studies regarding two
novel aspects. (a) In contrast to previous works,
our dataset contains categorical features that require
conversion into numerical features, a task achieved
through the implementation of the one-hot-encoding
method. However, this process substantially aug-
ments the dataset’s dimensionality, which poses a
challenge given the current constraints of noisy in-
termediate scale quantum (NISQ) computers. To ad-
dress this issue, we employ autoencoders (Bank et al.,
2020) to acquire a condensed representation of the
feature space that retains the essential information
from the initial data, as it was also done in the con-
text of anomaly detection with QML in (Wo´zniak
et al., 2023). (b) We extend the latter approach by
not only using fidelity quantum kernels but also pro-
jected quantum kernels (Huang et al., 2021), which
are promising candidates for the utility of near-term
quantum computing schemes. We finally analyze the
results of the different (Q)SVM methods and compare
the results of the encoded data sets with those of the
original data sets to investigate the impact of autoen-
coders on the entire pipeline.
The work is structured as follows: In section 2, we
present related work. Section 3 describes the method-
ology. The main findings are presented in Section 4
followed by a conclusion.
2 RELATED WORK
2.1 Machine Learning for Price
Prediction
In order to interpret the results within this paper and
to embed these findings in the current literature, let
us briefly recap the most related work and their spe-
cial focus. (Zong, 2017) estimates the residual value
of used articulated trucks using various regression
models. Similarly, (Chiteri, 2018) analyses the resid-
ual value of
3
4
ton trucks based on historical data
from auctions and resale transactions. Furthermore,
(Miloševi
´
c et al., 2021) construct an ensemble model
based on a diverse set of regression models to pre-
dict the residual value of 500 000 construction ma-
chines advertised in the USA. In addition, (Shehadeh
et al., 2021) and (Alshboul et al., 2021) use vari-
ous regression models to predict the residual value of
six construction equipment types based on data from
open-accessed auction databases. Finally, (Stühler.
et al., 2023) compared seven different state-of-the-art
and well-established ML methods with three AutoML
methods on a dataset generated from real online ad-
vertisements, consisting of 2910 datapoints from 10
different Caterpillar models.
All these research activities underline the advan-
tages and necessity of ML methods when dealing with
price forecasting applications.
2.2 Quantum Machine Learning
In the search for an advantage over classical meth-
ods with quantum computing, machine learning is ex-
pected to be one of the first fields to benefit from
quantum computers (Biamonte et al., 2017). Quan-
tum machine learning deals with incorporating quan-
tum algorithms for learning problems. Proven quan-
tum advantages in QML are based on algorithms that
can only be executed on fault-tolerant QCs (Shor,
1999; Harrow et al., 2009; Huang et al., 2021; Liu
et al., 2021). As this field just started with the emer-
gence of commercially available NISQ computers,
the practical implementation of the already known al-
gorithms is still in its infancy. There are three rea-
sons why we decided to use quantum support vec-
tor machines (QSVMs): (a) The proven speed-up in
a constructed theoretical problem based on the dis-
crete logarithm (Liu et al., 2021). (b) From a mathe-
matical point of view, classical SVMs are well under-
stood within the statistical learning theory in terms of
error bounds, convergence, robustness, and computa-
tional complexity (Steinwart and Christmann, 2008;
Schölkopf et al., 2002; Vapnik, 1999; Cortes and
Vapnik, 1995). Since for most QSVMs only the
kernel estimation is done on a quantum computer,
there are rigorous error bounds as well (Huang et
al., 2021). (c) QSVMs are especially suitable for the
NISQ era because of their shallow circuits. It should
be noted that the SVM optimization can be formu-
lated as a quadratic unconstrained binary optimiza-
tion (QUBO) problem (Willsch et al., 2020; Cavallaro
et al., 2020) and can hence be solved on a quantum
annealer (Kadowaki and Nishimori, 1998; Das and
Chakrabarti, 2005; Hauke et al., 2020). While this
is its own research branch, we focus on kernel-based
quantum regression in the scope of this paper.
Evaluating Quantum Support Vector Regression Methods for Price Forecasting Applications
377
Clear dataset of duplicates
Remove outliers
Generate dataset with different
feature combinations
Evaluate and document results
for all four feature combination
Train Models
One-hot encoding & scaling
Encoded
Unencoded
QSVR
!"#$
"%#$
QSVR
!"#$
"%#&'%
QSVR
!"#(
"%#&'%
QSVR
!"#(
"%#&'%
QSVR
!"#$
"%#$
QSVR
!"#$
"%#&'%
QSVR
)*+,
QSVR
)*+,
SVR
SVR
Figure 1: Methodology framework: The case study pipeline
illustrates the steps of the data processing phase in green
and the (Q)SVR algorithms in yellow. QSVR
en=0
re=0
is with-
out entanglement & data re-uploading. QSVR
en=lin
re=0
is with
linear entanglement & no data re-uploading. QSVR
en=lin
re=1
is with entanglement & one data re-uploading.
3 METHODOLOGY
To examine the QML capabilities, we compare four
quantum support vector regression (QSVR) methods,
which will be introduced in chapter 3.4 and chap-
ter 3.5 with a state-of-the-art classical support vec-
tor regression (SVR) implementation, introduced in
chapter 3.3. The overall method and the case study
pipeline are depicted in Figure 1 and explained in the
subsequent sections. The developed source code is
available on GitHub
1
.
3.1 Data Creation
The initial data was obtained by regularly collect-
ing all advertisements from seven major construction
equipment market portals
2
over a time period of seven
months. Table 1 shows the collected and selected fea-
tures.
1
See https://tinyurl.com/ymmts5xc.
2
The market portals are Mascus, Catused, Mobile, Ma-
chineryLine, TradeMachines, Truck1, and Truckscout24.
Figure 2: Working hours versus price for the Caterpillar 329
dataset.
Duplicate entries are eliminated by an iterative
comparison of different feature combinations. Out-
liers were detected by a plausibility check, namely
removing values outside a 99% confidence interval,
considering working hours and price.
Dealing with missing values depends on the at-
tribute. Samples are dropped if a value of the fea-
tures model, construction year, extension, or location
is missing. Missing values for the working hours at-
tribute will be substituted via stochastic regression
imputation (Newman, 2014). The entries for brand
and price are mandatory on all portals for creating ad-
vertisements.
In contrast to (Stühler. et al., 2023) for computa-
tional reasons, we only took the dataset of the Cater-
pillar model 329 with 141 data points. The distribu-
tion of the data points concerning the working hours
and price features are depicted in Figure 2. As we
took one machine model manufactured by Caterpillar,
the brand and the model feature, depicted in Table 1,
are thus obsolete. Table 2 shows an excerpt of the re-
sulting dataset. Data subsets with individual feature
combinations are created to account for and investi-
gate the impact of single features. The subset consist-
ing of the working hours and construction year is used
as the baseline feature set (subsequently referred to as
basic subset). In addition, this basic subset was ex-
Table 1: Collected dataset features with types and examples.
Feature Type Example
Brand Categorical Caterpillar
Model Categorical 329
Extension Categorical E
Construction year Numerical 2018
Working hours Numerical 8536
Location Categorical Germany
Price Numerical 59.000 C
ICAART 2024 - 16th International Conference on Agents and Artificial Intelligence
378
Table 2: Excerpt of the Caterpillar 329 dataset.
Extension Construction year Working hours [h] Location Price [C]
E 2012 10600 DE 77.000 C
D 2008 18180 CH 47.499 C
E 2012 11424 DE 72.900 C
E 2014 11500 DE 89.900 C
Table 3: Feature set combinations with the corresponding input and latent space dimensions. The features working hours &
constructin year form the basic subset and are the basis of all feature set combinations. The maximum size of the latent space
is 10. All feature sets are encoded to guarantee comparability.
Feature set Input size Latent size
basic subset 2
Autoencoded
2
basic subset + extension 4
Autoencoded
4
basic subset + location 16
Autoencoded
10
basic subset + extension + location 18
Autoencoded
10
tended by the extension and the location features and
all combinations of them, resulting in four data sets.
3.2 Autoencoding
The autoencoder (AE) (Tschannen et al., 2018; Ng
et al., 2011) is an unsupervised learning algorithm
that consists of two neural networks - the encoder and
the decoder (see Figure 3). While the former is used
to encode the input data in a reduced or dense rep-
resentation, called latent representation, the latter is
used to decode the original input from this reduced
representation. One advantage of using an AE is its
flexibility, i.e., the dimensionality of the latent space
representation can be changed by adding or removing
neurons, which is used to generate latent spaces de-
pending on the size or dimensionality of the datasets.
As mentioned in section 3.1, we are dealing with
four datasets with different, even feature set combi-
nations, which are displayed in Table 3 together with
their original and latent space dimensions. The fea-
tures working hours & construction year form the ba-
sic subset. The maximum size of the latent space is
10. Therefore, if the input dimension is greater than
10, it will be reduced to 10 by the AE. Feature sets
that are less than or equal to 10 are autoencoded with
the latent space size equal to the input space dimen-
sion. This is done to guarantee consistency for all
feature set combinations. We use the Adam optimizer,
the Mean-Squarer-Error loss, the Relu activation for
encoding, and the Sigmoid activation for decoding.
3.3 Support-Vector-Machines
Support vector machines (SVMs) are heavily used
ML methods for linear or nonlinear classification and
regression tasks mainly for small or medium-sized
datasets (Géron, 2022). In contrast to a "simple" lin-
ear classifier, a support vector classifier (SVC) tries
to find a separating hyperplane between two classes
such that the distance between the closest training in-
stances is maximized. This is called a large margin
classifier. These closest instances are called support
vectors. Hence the name support vector machine.
SVMs are designed to work in linearly separable
feature spaces. If the data points are not linearly sep-
arable in the original feature space, the feature space
can be transformed into a higher dimensional feature
space, where the problem becomes linearly separa-
Encoder Decoder
Latent Space
x
1
x
2
x
3
x
4
x
5
x
6
x
7
x
8
Input Layer
ˆx
1
ˆx
2
ˆx
3
ˆx
4
ˆx
5
ˆx
6
ˆx
7
ˆx
8
Output Layer
Hidden Layers Hidden Layers
Reduced Representation
Figure 3: A typical autoencoder consists of two deep neural
networks with multiple dense layers. The encoder trans-
forms the input data into the latent space representation,
which is reconstructed by the decoder. The autoencoder is
trained to reduce the reconstruction error; hence, it trains a
lower-dimensional representation of the input data in an ab-
stract mathematical space.
Evaluating Quantum Support Vector Regression Methods for Price Forecasting Applications
379
ble. This computationally expensive transformation
can be avoided using the kernel trick or kernel func-
tion. Commonly used kernel functions are the lin-
ear kernel, the polynomial kernel, and the radial ba-
sis function (RBF) kernel. Applying the kernel func-
tion to all data points results in the kernel matrix, a
square matrix containing the similarity measures be-
tween all data points in the dataset. The kernel ma-
trix is used to optimize the hyperplane that maxi-
mally separates the classes in the transformed, poten-
tially high-dimensional feature space. This, in turn,
avoids the computationally expensive transformation.
We use GridSearch for the classical SVR algorithm to
find the optimal kernel and hyperparameters.
3.4 Feature Maps
Quantum feature maps are a way of mapping clas-
sical data into a quantum state, which is then used
as an input to estimate a quantum kernel. There are
several ways to encode the data, including ampli-
tude encoding and the basis embedding method. The
choice of which method to use depends on the prob-
lem being addressed. In (Schuld et al., 2017; Havlí
ˇ
cek
et al., 2019), it was shown that amplitude encoding
outperformed the basis embedding method for cer-
tain datasets and vice versa for others. However, for
NISQ applications, we encode the classical data with
Pauli rotations as they comprise the most hardware-
efficient method. We encode the k-th feature of the
i-th data point x
k
i
[1,1] by using single-qubit rota-
tions around the x-, y- or z-axis on the Bloch-sphere
denoted by R
X
,R
Y
,R
Z
R
X
x
k
i
= e
iX x
k
i
/2
, (1)
R
Y
x
k
i
= e
iY x
k
i
/2
, (2)
R
Z
x
k
i
= e
iZ x
k
i
/2
, (3)
where X = |01|+ |10|, Y = i |01| + i |10|
and Z = |00||11| are Pauli matrices. A popular
basis gate set used on NISQ computers is
n
CNOT,Id,R
Z
,
X, X
o
, (4)
where CNOT is the only multi-qubit gate in that set
that is able to introduce entanglement into the feature
map. An example of such an encoding can be found in
Figure 4 (note that this is the untranspiled circuit). In
that feature map, we encode two features per qubit as
a good tradeoff between resources and the expressiv-
ity of the model. The quantum feature map should be
designed such that it cannot be simulated in polyno-
mial time on a probabilistic classical computer, in ac-
cordance with the Gottesman-Knill theorem (Aaron-
son and Gottesman, 2004).
q
0
:
H
R
Z
(πx
0
i
)
R
Y
(πx
n1
i
)
.
.
.
.
.
.
.
.
.
q
n1
:
H
R
Z
(πx
n1
i
)
R
Y
(πx
0
i
)
Figure 4: Quantum Feature Map for n features without any
entanglement (below: QSVR
en=0
re=0
)
q
0
:
R
X
(x
0
i
)
R
Y
(x
n1
i
)
q
1
:
R
X
(x
1
i
)
R
Y
(x
n2
i
)
q
2
:
R
X
(x
2
i
)
R
Y
(x
n3
i
)
.
.
.
.
.
.
.
.
.
q
n1
:
R
X
(x
n1
i
)
R
Y
(x
0
i
)
Figure 5: Quantum Feature Map for n features with an en-
tanglement between the encoding unitaries (as required by
the inversion test, below: QSVR
en=lin
re=0
).
Entanglement, which is applied on QSVR
en=lin
re=0
in
Figure 5, can enhance the expressive power of quan-
tum feature maps, but it can also make the implemen-
tations of QSVMs more difficult due to the increased
complexity of entangled quantum states. However,
another way of increasing the expressivity of a quan-
tum feature map is by the so-called data-reuploading
scheme (Schuld et al., 2021; Jerbi et al., 2023), which
is applied on QSVR
en=lin
re=1
in Figure 6 to make a more
expressive feature map. The entanglement layer,
which is implemented by a ring of CNOTs, is in be-
tween the encoding rotations due to the way quantum
kernel estimation is implemented (via the inversion
test) in some of our experiments.
3.5 Quantum Kernels
We use two types of quantum kernels in our exper-
iments: fidelity and projected quantum kernels. We
use the inversion test for the quantum kernel estima-
tion of the fidelity quantum kernels with the three fea-
ture maps introduced in section 3.4, whereas, for the
projected quantum kernels, we use the DensityMatrix
class from the Qiskit Quantum Information library.
Fidelity quantum kernels k(x
i
,x
j
) via the inversion
test are given by
k(x
i
,x
j
) =
ψ(x
i
)|ψ(x
j
)
2
, (5)
where ψ(x
i
) = U
φ
(x
i
)|0 . .. 0 is the quantum state
after passing through the quantum feature map
φ(x
i
) applied on the i-th classical datapoint. We
ICAART 2024 - 16th International Conference on Agents and Artificial Intelligence
380
q
0
:
R
X
(x
0
i
)
R
Y
(x
n1
i
)
R
X
(x
0
i
)
R
Y
(x
n1
i
)
q
1
:
R
X
(x
1
i
)
R
Y
(x
n2
i
)
R
X
(x
1
i
)
R
Y
(x
n2
i
)
q
2
:
R
X
(x
2
i
)
R
Y
(x
n3
i
)
R
X
(x
2
i
)
R
Y
(x
n3
i
)
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
q
n1
:
R
X
(x
n1
i
)
R
Y
(x
0
i
)
R
X
(x
n1
i
)
R
Y
(x
0
i
)
Figure 6: Quantum Feature Map for n features with entanglement and data re-uploading (below: QSVR
en=lin
re=1
).
rewrite Equation (5)
k(x
i
,x
j
) =
0.. .0|U
φ
(x
i
)U
φ
(x
j
)|0 . .. 0
2
,
=
0.. .0|Φ
i j
2
,
(6)
where we defined |Φ
i j
= U
φ
(x
i
)U
φ
(x
j
)|0 . .. 0.
Note that the term
0.. .0|Φ
i j
2
is just the prob-
ability of measuring |Φ
i j
in the |0... 0 state de-
noted by p
i j
(|0.. .0). Hence, the fidelity quan-
tum kernel is just the frequency of occurrences n
g
i j
of |Φ in the ground state
k(x
i
,x
j
) = p
i j
(|0.. .0) ,
=
n
g
i j
#shots
.
(7)
Using the quantum Hilbert space as the feature
space has the advantage that it grows exponen-
tially with the number of qubits used, which in
turn allows to obtain high-dimensional feature
maps. For instance, they are especially power-
ful for the creation of separating hyperplanes for
the classification of astronomy data (Peters et al.,
2021). However, a feature map that is too high-
dimensional for a specific learning task might
be too expressive and fails. This is known as
the curse of dimensionality. For this reason, we
also use projected quantum kernels (Huang et al.,
2021) because they circumvent this issue by pro-
jecting back into the classical space.
Projected Quantum Kernels: k(x
i
,x
j
) that we use
are given by
k(x
i
,x
j
) = exp
γ
n
q
k=0
ρ
k
(x
i
) ρ
k
(x
j
)
2
!
,
(8)
where γ > 0 and the reduced density matrix ρ
k
(x
i
)
denotes the density matrix ρ(x
i
) created by the
quantum feature map applied on the datapoint x
i
,
q
0
:
R
X
(x
0
i
) R
Y
(x
1
i
)
q
1
:
R
X
(x
2
i
) R
Y
(x
3
i
)
q
2
:
R
X
(x
4
i
) R
Y
(x
5
i
)
.
.
.
.
.
.
.
.
.
q
n1
:
R
X
(x
n2
i
) R
Y
(x
n1
i
)
Figure 7: Quantum Feature Map for n features with entan-
glement by CNOT gates (below: QSVR
proj
).
where all but the k-th qubit (of total n
q
) are traced
out. Note that it is sufficient for this type of quan-
tum kernel to know only the reduced density ma-
trix - not the full density matrix. We obtain these
by calculating the full density matrix and evalu-
ating the reduced density matrices by taking the
partial traces from it.
3.6 Performance Metric
Within this work, we use the accuracy to benchmark
the performance of the introduced methods.
Accuracy: measures the predictive power of an ML
model. In the context of this work, the mean ab-
solute percentage error (MAPE)
MAPE =
1
n
n
i=1
|y
i
ˆy
i
|
|y
i
|
(9)
is used to calculate the performance, with y
i
be-
ing the true value, ˆy
i
the predicted value and n the
number of data points.
4 RESULTS
This section presents the results of the experiments.
All measurements were performed on a Ubuntu Linux
20.04.5 LTS system with 64 GB RAM and an AMD
Ryzen Threadripper 192X 12-Core Processor. We
used Qiskit (IBM-Qiskit, 2023; IBM, 2018) and
Evaluating Quantum Support Vector Regression Methods for Price Forecasting Applications
381
Figure 8: Accuracy for the encoded feature sets, in form of MAPE. Diamonds depict the average results over 5 repetitions.
Single measurements are displayed as dots. QSVR
proj
needs at least four features; therefore, there are no basic subset results
for QSVR
proj
.
Table 4: Overview of the MAPE for the encoded and unencoded datasets of the complete feature set. The best results are
highlighted in bold. * SVR and QSVR
proj
produce identical results for the unencoded, identical datasets.
Method Encoded datasets [MAPE] Unencoded datasets [MAPE]
SVR 0.1560±0.0133 0.1452 *
QSVR
en=0
re=0
0.1349±0.0110 0.1517±0.0025
QSVR
en=lin
re=0
0.1521±0.0181 0.1368±0.0079
QSVR
en=lin
re=1
0.1380±0.0102 0.1481±0.0076
QSVR
proj
0.1695±0.0216 0.1511 *
Qiskit Aer (Wood, 2019) for all QC simulations and
conducted five independent measurements with the
identical 80% / 20% holdout training/test split.
Accuracy
The accuracy, in terms of the MAPE, of the differ-
ent methods is depicted in Figure 8 for the encoded
datasets and in Figure 9 for the unencoded datasets. In
particular, these figures visualize the following char-
acteristics:
1 On the encoded datasets, all of the applied QSVR
methods, with the exception of QSVR
proj
for the
complete dataset, perform better than the classical
SVR implementation.
2 On the unencoded datasets, with the ex-
ception of the basic dataset, QSVR
en=lin
re=0
or
QSVR
proj
perform better than the classical SVR
method.
3 All results indicate good prediction qualities for
the encoded and unencoded datasets as they are
roughly within the interval [0.13, 0.17].
4 QSVR
en=0
re=0
performs best in three out of four en-
coded feature sets. It shows the overall best per-
formance for the encoded feature set but worst
in three out of four unencoded feature sets.
QSVR
en=lin
re=0
performs best for the unencoded fea-
ture sets. Examining these results in detail and the
impact of AE on the different QSVR methods will
be systematically analyzed in future work.
5 The encoded datasets exhibit much more variance
than the unencoded datasets. This is due to the
variance of the autoencoder. This exhibits the in-
fluence of the AE on the accuracy of the results.
This fact needs to be taken into account for each
calculation.
6 The impact of the features is obvious for both un-
encoded and encoded datasets. Adding the ex-
tension features to the basic subset downgrades
the accuracy within the unencoded and encoded
datasets. The best result is always achieved within
ICAART 2024 - 16th International Conference on Agents and Artificial Intelligence
382
Figure 9: Accuracy for the unencoded feature sets, in form of MAPE. Diamonds depict the average results over 5 repetitions.
Single measurements are displayed as dots. QSVR
proj
needs at least four features; therefore, there are no basic subset results
for QSVR
proj
.
the complete feature set.
The MAPE values of all five methods for the complete
feature set of the encoded and unencoded data sets are
depicted in Table 4.
5 CONCLUSION
This work analyzed the potential of QMLs methods as
a substitution for linear, polynomial, or RBF kernels
for the SVR algorithm. In our case study, predicting
the residual value of used heavy construction equip-
ment, all simulated QML methods were shown to be
comparable or even better than the classical SVM
kernels regarding the MAPE scores. Therefore, we
showed that current state-of-the-art QSVR algorithms
are able to substitute the classical SVR implementa-
tions.
It has to be mentioned that we only examined a
limited number of QML methods on four variations of
a single data set, so general statements are therefore
limited by our choice of methods.
In order to strengthen the significance of our
statistics, we will extend our survey to additional
and in particular open-source data sets. This analy-
sis will enable us to validate and substantially gen-
eralize the statements of our paper. Furthermore, we
will examine the role of entanglement and the power
of quantum kernels in more detail. We will analyze
the impact of dimensionality reduction, which will be
mandatory in the near future for datasets with larger
feature sets, on the quantum model performance in
future work. Finally, we will add more QML meth-
ods to our case study and integrate a real QC hard-
ware backend into our framework to be able to run
the QML algorithms on a real QC.
ACKNOWLEDGEMENTS
This work was partly funded by the German Federal
Ministry of Economic Affairs and Climate Action in
the research project AutoQML.
REFERENCES
Aaronson, S. and Gottesman, D. (2004). Improved sim-
ulation of stabilizer circuits. Physical Review A,
70(5):052328.
Alshboul, O., Shehadeh, A., Al-Kasasbeh, M., Al Mam-
look, R. E., Halalsheh, N., and Alkasasbeh, M. (2021).
Deep and machine learning approaches for forecasting
the residual value of heavy construction equipment:
a management decision support model. Engineering,
Construction and Architectural Management.
Bank, D., Koenigstein, N., and Giryes, R. (2020). Autoen-
coders. arXiv preprint arXiv:2003.05991.
Biamonte, J., Wittek, P., Pancotti, N., Rebentrost, P., Wiebe,
N., and Lloyd, S. (2017). Quantum machine learning.
Nature, 549(7671):195–202.
Cavallaro, G., Willsch, D., Willsch, M., Michielsen, K.,
and Riedel, M. (2020). Approaching remote sensing
Evaluating Quantum Support Vector Regression Methods for Price Forecasting Applications
383
image classification with ensembles of support vector
machines on the d-wave quantum annealer. In IGARSS
2020-2020 IEEE International Geoscience and Re-
mote Sensing Symposium, pages 1973–1976. IEEE.
Chiteri, M. (2018). Cash-flow and residual value analysis
for construction equipment. Master’s thesis, Univer-
sity of Alberta.
Cortes, C. and Vapnik, V. (1995). Support-vector networks.
Machine learning, 20:273–297.
Das, A. and Chakrabarti, B. K. (2005). Quantum anneal-
ing and related optimization methods, volume 679.
Springer Science & Business Media.
Géron, A. (2022). Hands-on machine learning with Scikit-
Learn, Keras, and TensorFlow. " O’Reilly Media".
Grossi, M., Ibrahim, N., Radescu, V., Loredo, R., Voigt, K.,
von Altrock, C., and Rudnik, A. (2022). Mixed quan-
tum–classical method for fraud detection with quan-
tum feature selection. IEEE Transactions on Quantum
Engineering, 3:1–12.
Harrow, A. W., Hassidim, A., and Lloyd, S. (2009). Quan-
tum algorithm for linear systems of equations. Physi-
cal review letters, 103(15):150502.
Hauke, P., Katzgraber, H. G., Lechner, W., Nishimori, H.,
and Oliver, W. D. (2020). Perspectives of quantum
annealing: Methods and implementations. Reports on
Progress in Physics, 83(5):054401.
Havlí
ˇ
cek, V., Córcoles, A. D., Temme, K., Harrow, A. W.,
Kandala, A., Chow, J. M., and Gambetta, J. M. (2019).
Supervised learning with quantum-enhanced feature
spaces. Nature, 567(7747):209–212.
Huang et al., H.-Y. (2021). Power of data in quantum ma-
chine learning. Nature Communications, 12(1):2631.
IBM (2018). An open high-performance simulator for
quantum circuits. IBM Research Editorial Staff.
IBM-Qiskit (2023). Qiskit github page. https://github.com/
Qiskit.
Jerbi, S., Fiderer, L. J., Nautrup, H. P., Kübler, J. M.,
Briegel, H. J., and Dunjko, V. (2023). Quantum ma-
chine learning beyond kernel methods. Nature Com-
munications, 14(1).
Kadowaki, T. and Nishimori, H. (1998). Quantum anneal-
ing in the transverse ising model. Physical Review E,
58(5):5355.
Liu, Y., Arunachalam, S., and Temme, K. (2021). A rig-
orous and robust quantum speed-up in supervised ma-
chine learning. Nature Physics, 17(9):1013–1017.
Lucko, G., Vorster, M. C., and Anderson-Cook, C. M.
(2007). Unknown element of owning costs - impact
of residual value. JCEMD4, 133(1).
Miloševi
´
c, I., Kova
ˇ
cevi
´
c, M., and Petronijevi
´
c, P. (2021).
Estimating residual value of heavy construction equip-
ment using ensemble learning. JCEMD4, 147(7).
Newman, D. A. (2014). Missing data: Five practical guide-
lines. Organizational Research Methods, 17(4).
Ng, A. et al. (2011). Sparse autoencoder. CS294A Lecture
notes, 72(2011):1–19.
Peters et al., E. (2021). Machine learning of high dimen-
sional data on a noisy quantum processor. npj Quan-
tum Information, 7(1):161.
Schölkopf, B., Smola, A. J., Bach, F., et al. (2002). Learn-
ing with kernels: support vector machines, regular-
ization, optimization, and beyond. MIT press.
Schuld, M., Fingerhuth, M., and Petruccione, F. (2017).
Implementing a distance-based classifier with a
quantum interference circuit. Europhysics Letters,
119(6):60002.
Schuld, M., Sweke, R., and Meyer, J. J. (2021). Effect of
data encoding on the expressive power of variational
quantum-machine-learning models. Physical Review
A, 103(3):032430.
Shehadeh, A., Alshboul, O., Al Mamlook, R. E., and Hame-
dat, O. (2021). Machine learning models for predict-
ing the residual value of heavy construction equip-
ment: An evaluation of modified decision tree, light-
gbm, and xgboost regression. Automation in Con-
struction, 129:103827.
Shor, P. W. (1999). Polynomial-time algorithms for prime
factorization and discrete logarithms on a quantum
computer. SIAM review, 41(2):303–332.
Steinwart, I. and Christmann, A. (2008). Support vector
machines. Springer Science & Business Media.
Stühler., H., Zöller., M., Klau., D., Beiderwellen-
Bedrikow., A., and Tutschku., C. (2023). Benchmark-
ing automated machine learning methods for price
forecasting applications. In Proceedings of the 12th
International Conference on Data Science, Technol-
ogy and Applications - DATA, pages 30–39. INSTICC,
SciTePress.
Thanasilp, S., Wang, S., Cerezo, M., and Holmes, Z.
(2022). Exponential concentration and untrainabil-
ity in quantum kernel methods. arXiv preprint
arXiv:2208.11060.
Tschannen, M., Bachem, O., and Lucic, M. (2018). Recent
advances in autoencoder-based representation learn-
ing. arXiv preprint arXiv:1812.05069.
Vapnik, V. N. (1999). An overview of statistical learn-
ing theory. IEEE transactions on neural networks,
10(5):988–999.
Willsch, D., Willsch, M., De Raedt, H., and Michielsen,
K. (2020). Support vector machines on the d-wave
quantum annealer. Computer physics communica-
tions, 248:107006.
Wood, C. J. (2019). Introducing qiskit aer: A high per-
formance simulator framework for quantum circuits.
Medium.
Wo´zniak, K. A., Belis, V., Puljak, E., Barkoutsos, P., Disser-
tori, G., Grossi, M., Pierini, M., Reiter, F., Tavernelli,
I., and Vallecorsa, S. (2023). Quantum anomaly de-
tection in the latent space of proton collision events at
the lhc. arXiv preprint arXiv:2301.10780.
Zöller, M.-A., Nguyen, T.-D., and Huber, M. F. (2021).
Incremental search space construction for machine
learning pipeline synthesis. In Advances in Intelligent
Data Analysis XIX.
Zong, Y. (2017). Maintenance cost and residual value pre-
diction of heavy construction equipment. Master’s
thesis, University of Alberta.
Zoph, B. and Le, Q. V. (2016). Neural architecture
search with reinforcement learning. arXiv preprint
arXiv:1611.01578.
ICAART 2024 - 16th International Conference on Agents and Artificial Intelligence
384