Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

International Journal of Machine Learning and Cybernetics

https://doi.org/10.1007/s13042-019-01036-y

ORIGINAL ARTICLE

A self‑adaptive preference model based on dynamic feature analysis


for interactive portfolio optimization
Shicheng Hu1 · Fang Li1 · Yang Liu2 · Song Wang1

Received: 19 February 2019 / Accepted: 5 November 2019


© Springer-Verlag GmbH Germany, part of Springer Nature 2019

Abstract
In financial markets, there are various assets to invest in. Recognizing an investor’s preferences is key to selecting a com-
bination of assets that best serves his or her needs. Considering the mean–variance model for the portfolio optimization
problem, this paper proposes an interactive multicriteria decision-making method and explores a self-adaptive preference
model based on dynamic feature analysis (denoted RFFS-DT) to capture the decision maker (DM)’s complex preferences in
the decision-making process. RFFS-DT recognizes the DM’s preference impact factor and constructs a preference model.
To recognize the impact factors of the DM’s preferences, which could change during the decision-making process, three
categories of possible features involved in three aspects of the mean–variance model are defined, and a feature selection
method based on random forest is designed. Because the DM’s preference structure could be unknown a priori, a decision-
tree-based preference model is built and updated adaptively according to the DM’s preference feedback and the selected
features. The effectiveness of RFFS-DT for interactive multicriteria decision making is verified by a series of deliberately
designed comparative experiments.

Keywords Feature selection · Random forest · Preference model · Portfolio optimization

1 Introduction decomposition) [4]. Unfortunately, the output of an EMO


consists of a large number of nondominated solutions, and
Investment is a fundamental economic activity in our soci- common users do not know how to choose the solution that
ety. Because there are many assets with different returns is in his or her best interest [5].
and risks in the financial market, investors have to choose For a multiobjective optimization problem, another group
assets to gain the highest return possible while incurring the of methods, multicriteria decision-making methods [6, 7],
lowest risk possible. To address this issue, Nobel laureate is more practical than the EMO. Multicriteria decision
Harry Markowitz [1, 2] introduced a portfolio optimization making aims to help a DM find the most preferred solution
model, the mean–variance (MV) model, to determine the from the numerous candidates produced by the EMO pro-
optimal allocation of limited capital among a finite num- cess. There are three types of multicriteria decision-making
ber of available assets. The MV model can be solveed by approaches: the a priori approach [8, 9], the a posteriori
an evolutionary multiobjective algorithm (EMO), e.g., approach [10, 11] and the interactive approach [12, 13]. For
NSGAII (nondominated sorting genetic algorithm II) [3] and these approaches, the DM’s preferences are incorporated
MOEA/D (multiobjective evolutionary algorithm based on before, after and during the EMO process, respectively.
However, for these approaches, the difficulty of obtaining
the preferences of the DM is not considered. The prefer-
* Shicheng Hu ence information given in the a priori approach and the a
hu_shicheng@aliyun.com
posteriori approach should be global, but in the interactive
1
School of Economics and Management, Harbin Institute approach, the preference information could be local, as it can
of Technology, No. 2 West Wenhua Road, Weihai 264209, be interactively adjusted by the DM. From the perspective of
China human cognition, an interactive approach is easier to imple-
2
School of Computer Science and Technology, Harbin ment and therefore more feasible for solving multiobjective
Institute of Technology, No. 2 West Wenhua Road, optimization problems.
Weihai 264209, China

13
Vol.:(0123456789)
International Journal of Machine Learning and Cybernetics

Interactive multicriteria decision making (IMCDM) In IMCDM, the preference model and preference feed-
[12–15] is composed of three parts: an EMO algorithm, the back are closely related. From the above analysis, we can
DM’s preference feedback and a preference model. Initially, conclude that utility-function-based IMCDM is practical
the EMO algorithm runs for a certain number of iterations. with regard to preference feedback; however, the func-
Then, the current nondominated solutions are presented to tional form of the preference model is difficult to choose in
the DM for preference feedback. Based on the DM’s feed- accordance with the DM’s complicated and variable prefer-
back, a preference model is built to guide the EMO algo- ences. In this paper, for portfolio optimization, we propose
rithm. This process is continued until the DM is satisfied a self-adaptive preference model based on dynamic feature
with the best of the presented solutions (i.e., the solution analysis. Our contributions, as they relate to IMCDM, are
with the highest rank, as evaluated by the preference model). described below.
Many interactive approaches have been proposed, and they First, the DM’s preference feedback is based on dynamic
can be grouped into three sets in terms of their preference feature analysis. In the IMCDMs proposed thus far, the
models: scalarizing-function-based IMCDM, utility-func- DM’s preference feedback is normally given based on the
tion-based IMCDM and neural-network-based IMCDM. objective functions of a multiobjective optimization model,
These approaches have investigated each aspect of IMCDM which means that the impact factor on the DM’s preference
and achieved fruitful results. However, considering their is predetermined (i.e., the objective function) and invariable.
overall performance, there are still deficiencies and chal- However, an optimization model is composed of objective
lenges to be addressed. functions, decision variables and constraints. These three
The scalarizing function S(y, p, w), which has a variety of components are directly mapped to each other by different
forms [16], scalarizes a multiobjective optimization problem formulas and hence intrinsically affect the DM’s preferences.
into a single objective optimization problem. The DM can For example, in a multiobjective TSP (traveling salesman
change his or her preferences directly either by specifying problem) [35], both the decision variable values that rep-
a different expectation p in the means of reference points resent a route and the satisfaction degree of the constraints
[17–20] and categories [21, 22] or by setting a different can affect the DM’s preferred itinerary. If the DM is always
weight vector w [23, 24]. Nevertheless, a precise quantitative confronted with objective function values, he/she may not
vector (p or w) is very difficult to obtain from a typical DM develop a deeper understanding of the problem during the
with bounded rationality [25, 26]. Utility-function-based decision-making process. As a result, the DM may be unable
IMCDM is very flexible because a wide range of functional to reasonably judge an interaction.
forms, such as linear [27], quadratic [28] and general poly- Second, for the preference model, we adopt a decision tree
nomial functions [25, 29], can meet the requirements of a adaptively built on dynamic feature analysis. The decision
DM with complicated preferences. Furthermore, the DM’s tree is completely constructed by an induction algorithm;
preference feedback can be determined through qualitative thus, the difficulty of preselecting a proper function form
pairwise comparison, which is much easier to implement for the utility function-based preference model is avoided.
than quantity-based preference feedback. However, in this During the decision process, if the DM’s preference struc-
type of IMCDM, the proper form of the utility function has ture varies without any prior knowledge, the decision tree
to be selected a priori, which is often impossible since the can be updated accordingly; furthermore, through dynamic
DM’s preference structure is unknown in most cases; the feature analysis, the decision tree’s components can be self-
neural network can simulate any form of the function, and adapted to the DM’s preference feedback. However, for the
its parameters can be learned adaptively during the decision- utility-function-based preference model, the function form
making process [30–32]. However, neural-network-based will not change once selected, and its components are always
IMCDM requires the DM to give a ratio value for each pair restricted to objective functions. Therefore, the self-adaptive
of solutions. This highly precise quantification of the prefer- decision-tree-based preference model is much more flexible
ence feedback is beyond the scope of a typical human’s cog- than the utility-function-based preference model, and it is
nitive capability [25, 26]. Recently, two dynamic preference more suitable for capturing the DM’s complicated prefer-
models based on IMCDMs with automatic interaction were ences in the decision-making process than other models.
proposed to solve optimization problems related to the self- The rest of the paper is organized as follows. The dynamic
organization of heterogeneous wireless networks and object- feature analysis for the portfolio optimization problem is
level video advertising [33, 34]. These methods have proven detailed in Sect. 2. A self-adaptive preference model built
to be very effective and efficient, but their automatic method on feature analysis is presented in Sect. 3. The overall frame-
of interaction is not suitable for portfolio optimization, in work for interactive portfolio optimization is described in
which the DM’s preference feedback is needed periodically Sect. 4. The experimental evaluation is reported in Sect. 5.
so that variable preferences can be captured. Finally, conclusions are drawn.

13
International Journal of Machine Learning and Cybernetics

2 Dynamic feature analysis for portfolio 2.1 Feature definition


optimization
When using the MV model for the portfolio optimization
Portfolio optimization can be used to solve an optimization problem, possible features can be defined in terms of three
problem involving the allocation of limited capital among aspects.
a certain number of assets (e.g., stocks) to gain the highest
return at the lowest risk. The most influential MV model (1) Objective functions
for portfolio optimization was introduced by Markowitz
[1, 2], and it can be stated as follows: The maximization of the return and the minimization of
risk are the two objectives of the MV model. The features
N
∑ that may be related to the objectives include the following:
max f1 = 𝜔i 𝜇i (1)
i=1
Maximum return. By investing, the DM hopes to gain the
highest possible return [36].
N N Minimum risk. Since the financial market is full of uncer-
tainty, investment may result in capital losses [37]. The min-
∑ ∑
min f2 = 𝜔i 𝜔j 𝜎ij (2)
i=1 j=1 imum risk minimizes any losses.
Expected return. As the MV is a multiobjective optimi-
N zation problem, for any two solutions in the nondominated
set, the one with highest return will have the highest risk

s.t. 𝜔i = 1 (3)
i=1 [38]. Considering his or her reason for investing, an inves-
tor may have an expectation for the return regardless of the
0 ≤ 𝜔i ≤ 1, i = 1, … , N (4) possible risk. The investor’s expectation may change dur-
ing the decision-making process. Thus, the margin between
where N is the number of assets available to a DM, 𝜇i the investor’s expectation and the return of a solution may
(i = 1, … , N ) is the expected return of asset i, 𝜎ij(i = 1, … , N impact his or her preferences.
; j = 1, … , N ) is the covariance between assets i and j, and Expected risk. Similar to the expected return, an investor
𝜔i is the decision variable indicating the proportion of lim- may also have a level of expected risk.
ited capital invested in asset i. Equation (1) maximizes the The distance to the ideal point. Although an ideal solu-
total returns associated with the portfolio, while Eq. (2) tion with the highest return and the lowest risk does not
minimizes the total risks. Equation (3) is the budget con- exist in the decision domain [39], a solution close to the
straint that requires limited capital to be invested in the ideal point balances the return and risk of the portfolio. The
available assets. Equation (4) requires all investments to be distance of this solution to the ideal point may impact the
nonnegative. DM’s preferences.
The MV model can be used to solve a two-objective
optimization problem. To obtain a solution that satisfies (2) Decision variables
a DM, an EMO must be implemented in an interactive
manner. Apparently, the preference model plays a very The decision variables represent the proportion of the
important role in the interactive decision process. Whether budget invested in each asset [38]. The related impact factors
the solution produced by the EMO satisfies the DM mainly can be defined as follows:
depends on whether the preference model includes the Proportion of each asset. In many cases, when the DM
DM’s true preference information. The DM’s preferences cannot select the preferred solution from a pair of solutions
may change during the interactive process. During each in terms of only return and risk, the decision variable values
interaction, the DM cannot reveal the impact factors of his can provide additional information to help the DM make a
or her preferences due to his or her bounded rationality. In decision.
this case, the DM’s preference factors can be determined Expected proportions of some assets. Regarding the
only from the preference feedback and then embedded expected return and risk, an investor may have a preference
into the preference model. The DM’s preference factors regarding the expected proportions of some assets, which
are determined in two steps: feature definition and feature can influence his or her preferences.
selection. In the first step, all features that can impact the
DM’s preferences are defined. In the second step, a sophis- (3) Constraint conditions
ticated feature selection method is designed to discern the
DM’s true preference factors. In the MV model, intrinsic constraints (3) and (4) are hard
constraints [40] that should be satisfied during the solution

13
International Journal of Machine Learning and Cybernetics

process. With respect to the DM’s uncertain and complicated Initialization


preferences, some additional constraints may be embedded
during the interactive decision-making process. These addi-
Create k random forests
tional constraints are soft constraints [40] that are not neces-
sarily justified.
Reservation value for return. The reservation value [41] Compute classification accuracy for
is the objective value that should be achieved. In investment, each random forest and mean accuracy
some risky investors desire a minimal return regardless of LMeanAcc
the possible risk.
Reservation value for risk. Some risk-averse investors Select the most accurate random forest
prefer risk that is below a certain level.
Reservation value for some assets. In the financial mar-
Sort the features in Data
ket, an experienced investor tends to select a portfolio that
includes familiar assets he/she has benefited from in the past.
N
The investor may have a preference regarding the minimum LMeanAcc > MaxMeanAcc ?
proportions that are invested in those assets.
The reservations can be set as soft constraints. The mar- Y
gin between the reservation values and real values of a port- Delete the least important
folio may impact the DM’s preferences. feature from Data

MaxMeanAcc = LMeanAcc
2.2 Feature selection method RFFS
N
Feature selection is applied after the DM articulates his or Number of iterations is reached?
her preference on a set of currently nondominated solutions
with defined features. To analyze the correlation between the Y
possible features and the DM’s preference feedback in the Output Data
decision-making process, we design a random-forest-based
feature selection method (RFFS). Random forest [42] is an Fig. 1  Flowchart of RFFS implementation
ensemble classifier composed of a group of weak learners
(e.g., decision trees). In most cases, random forest exhibits
a high classification accuracy and a short training time [43]; is set as the number of features in Data. The output of the
furthermore, due to its randomness, it is not very sensitive RFFS is Data with the least important features removed.
to noise [44].
In the RFFS, the mean decrease in classification accu-
racy of the out-of-bag data [45] of a bootstrap sample is 3 Self‑adaptive preference model RFFS‑DT
adopted as the feature/variable importance evaluation cri-
terion; sequential backward selection [46] is employed as The preference model is built on a set of currently nondomi-
the feature search strategy, and 10-CV (cross-validation) is nated solutions with a subset of truncated features by the
used for validation. RFFS, in which the DM’s preference feedback is used for
The implementation of the RFFS is presented in Fig. 1. class labels. Since the truncated feature subset can be var-
For the implementation of the RFFS, the following vari- ied during the decision-making process, the function-based
ables should be initialized: the current nondominated solu- model is not suitable due to its preselected function form
tion set Data with all defined features; the number of random with fixed variables/features. Given a dataset, a classifier
forests created in each iteration (k = 10 since 10-CV is used can be induced by a learning algorithm without the need for
for validation); and the incumbent maximum mean classi- a preselected structure form, which means that any classifier
fication accuracy for k random forests (MaxMeanAcc = 0). can be used as the preference model for the DM’s variable
In each iteration, k random forests are created on Data. To preference features. However, many opaque classifiers, such
compute their classification accuracies and the mean accu- as neural networks and Bayesian classifiers, cannot give a
racy, LMeanAcc, Data is partitioned into k groups, and each clear interpretation of their classification results. An ensem-
is used as a test set. The features in Data are sorted in terms ble classifier, even though its members may be transparent,
of the variable importance evaluation criterion by the most produces output based on a majority vote among its mem-
accurate random forest selected. The number of iterations bers and hence provides no explanation of its classification.

13
International Journal of Machine Learning and Cybernetics

The decision tree [47] is a sequential model; each path establishes a one-to-one map between a local Pareto front
from a root node to a leaf node represents a classification of a multiobjective optimization problem and a unified sam-
rule that closely resembles human reasoning and is thus easy ple space 𝛺.
to understand. Due to its superiority in terms of comprehen- In the MV model, we change objective (1) so that both
sibility a decision tree would make the DM feel more com- objectives (1) and (2) are minimizations, as given in (6) and
fortable than a black-box model. Furthermore, as the DM’s (7), respectively, and then apply (5) to them both.
understanding of the problem improves by learning the deci- N
sion tree’s classification rules, his or her preference feedback ∑
min z1 = 1 − f1 = 1 − 𝜔i 𝜇i (6)
may become more confident. As a result, the convergence of i=1
the interactive decision process can be accelerated.
The decision tree (DT)-based preference model is com- N N
pletely self-adapted to the DM’s preference feedback. At
∑ ∑
min z2 = f2 = 𝜔i 𝜔j 𝜎ij (7)
each interaction, the DM is provided a set of nondominated i=1 j=1
solutions with a full set of features. In terms of the DM’s
preference feedback, the RFFS is applied to reduce unim- As the return f1 in the MV model is less than one, z1 is
portant features. The decision tree model then learns using always positive, which avoids having a value of zero in the
the same set of solutions with a subset of truncated features. denominator of Eq. (5).
Without any assumptions, the generation of the DT-based
preference model during the interactive decision-making 3.2 Preference model’s guidance for the EMO
process is dynamic and automatic. In terms of its inter-
pretability, the self-adaptive DT-based preference model The objective of EMO is to produce a set of optimal and
can fulfill the needs of a DM with uncertain and flexible representative solutions, and nondominance and diversity
preferences. [52] are used to order the generated solutions. However,
To date, many decision tree induction algorithms have the ultimate goal of IMCDM is to search for a solution that
been proposed [48, 49]. A CART (classification and regres- satisfies the DM [25, 28]. A satisfactory solution should be
sion tree) is employed in our IMCDM due to its simplicity optimal and reflect the DM’s preferences; thus, the generated
of implementation, high efficiency, efficacy and wide range solutions should be evaluated by a nondominance model and
of successful applications [50, 51]. However, to use the MV a preference model. Hence, in the EMO process of IMCDM,
model for portfolio optimization, some issues related to the the diversity is replaced by the preference model.
DT-based preference model must be addressed. In our IMCDM for the MV model, nondominance and
preference model-based ordering is applied in two parts of
the EMO:
3.1 Establishment of a unified sample space Production of child individuals. Child individuals are pro-
duced from parent individuals. To produce better children,
As a classifier, a decision tree implicitly assumes that the elitist parents are chosen according to nondominance and
datasets to be processed are sampled from a space with com- preference model-based ordering.
mon distribution characteristics. However, the DT-based Production of the child population. Not all the child
preference model in our IMCDM is built on a set of cur- individuals produced are better than their parent individu-
rently nondominated solutions. The different sets of non- als; thus, the two groups are merged and then ordered by
dominated solutions during the iteration process are pro- the nondominance and preference model to produce better
duced from local Pareto fronts with different distribution individuals.
characteristics. Therefore, they should be mapped into one
space with common distribution characteristics so that the
DT-based preference model is built therein. For an optimi- 4 Framework for interactive portfolio
zation problem with M objectives, zl (l = 1, 2, …, M), the optimization
nondominated solutions distributed on one Pareto front bal-
ance the objectives differently and can be transformed into The framework of our IMCDM for the MV model is
as represented by ratios using the following equation: depicted in Fig. 2. This framework is mainly composed of
z five modules: the EMO, preference feedback, feature selec-
𝜁l = ∑ l (5) tion, preference model creation/update and model evalua-
l zl
tion. The input includes an instance of the MV model (by
As l 𝜁l = 1, all the points in space 𝛺= { l 𝜁l = 1} instantiation) and the initial parameter values of IMCDM (by
∑ ∑
observe one characteristic distribution. Equation (5) initialization): s, the population size; max_t, the maximum

13
International Journal of Machine Learning and Cybernetics

Instantiation and initialization 5 Experimental evaluations

Experimental evaluation aims to verify the effectiveness of


Run plain EMO our IMCDM for the MV model by comparing the proposed
preference model RFFS-DT and the function-based prefer-
DM’s preference feedback ence model adopted in the literature. The examples used for
the MV model are generated from the well-known RESSET
financial research database, which includes a wide range of
Model evaluation data that are frequently cited for research on subjects such as
economics, finance, accounting, statistics and mathematics.
RESSET includes the latest data on daily to yearly scales,
RFFS based feature selection and these data can be openly accessed (http://www.resse​
t.cn). For 2017, there are 2849 stocks in 18 categories. We
Create/update compute the average monthly return for each stock as its
DT-based preference model expected return and randomly select one stock from each
category as an asset for the MV model. The 18 selected
stocks and their returns are listed in the appendix.
Run improved EMO
5.1 Comparative experimental design

Terminal condition is met? As mentioned in the introduction, a normal IMCDM


includes an EMO, preference feedback and a preference
model. To verify the effectiveness of our proposed RFFS-
End DT, IMCDM is run for the same EMO and preference feed-
back but different preference models.
Many EMOs can be used for the MV model. Among
Fig. 2  Framework of our IMCDM for the MV model
them, NSGAII [3] and MOEA/D [4] are the most influential.
However, recent studies have shown that MOEA/D outper-
number of generations; max_it, the maximum number of forms NSGAII both in terms of diversity and convergence
interactions; gen (= max_t/max_it), the number of itera- [29]; hence, MOEA/D is adopted.
tions between two interactions; st, the number of solutions Following the literature on IMCDM, we use functions to
presented to the DM for preference feedback. After instan- mimic DM’s true preferences, by which the DM’s pairwise
tiation and initialization, a plain EMO with nondominance comparison-based preference feedback is given. To validate
and diversity-based ordering and selection runs for the the comparative results, these functions are designed using
first gen iterations. After the first interaction, the DT-based simple and complex formulas as follows:
preference model is created, and then the improved EMO
with nondominance and preference model-based ordering V(f1 , f2 ) = 10f13 + 0.06f2 (8)
and selection runs for the next iterations. The two rankings
for the same set of solutions produced by the preference V(f1 , f2 ) = e3.1f1 + 2.64f2 (9)
model and the DM’s preference feedback are used for model
evaluation.
The DM can articulate his or her preferences in several
V(f1 , f2 ) = − ln(f1 ) + 15.625f2 . (10)
ways, such as evaluating each solution [30–32], ranking the These functions assume that only the objective functions
solutions [53–55] or conducting a pairwise comparison [25, of the optimization problem influence the DM’s preferences.
28, 29]. The qualitative pairwise comparison is the easiest To mimic the impact of additional factors on the DM’s pref-
to implement for a DM with bounded rationality [26, 56]; erences, we derive another set of functions by appending a
hence, this method is adopted to obtain the DM’s preference term to the above three functions. We presume the additional
feedback. Meanwhile, Saaty’s AHP (analytical hierarchy impact factor to be the decision variable x1 of asset no 1. The
process) [57] is used to rank the solutions. derived preference functions are given below:
The terminal condition is either achieving DM satisfac-
tion (the best solution, i.e., the highest ranked solution) or V(f1 , f2 , x1 ) = 10f13 + 0.06f2 + 0.05x1 (11)
reaching the maximum number of iterations (max_t) or inter-
actions (max_it).

13
International Journal of Machine Learning and Cybernetics

V(f1 , f2 , x1 ) = e3.1f1 + 2.64f2 + 0.05x1 (12) methods, it inherits the same value for st. In this paper, we
study the effect of two other parameters, max_it and gen, on
our method and determine their optimal values.
V(f1 , f2 , x1 ) = − ln(f1 ) + 15.625f2 + 0.05x1 . (13)
We compare RFFS-DT with the function-based prefer- 5.2 Comparative results
ence models in two cases: a DM with a stable preference
during the decision-making process and a DM with a vari-
able preference during the decision-making process. In the (1) DM with a stable preference
first case, the DM’s preference feedback is given by func-
tions (8)–(13). In the second case, the DM’s preference Tables 1, 2, and 3 show the results of a comparison of
feedback is given by functions (8)–(10) and by functions three preference models; the DM’s preference feedback is
(11)–(13) in different stages of interaction. The two most given based on the polynomial function (8), exponential
cited function-based preference models in the literature are function (9) and logarithmic function (10).
employed for a comparison; these functions are the linear From Table 1, we can see that G-PM has the minimum
function [27] and the general polynomial function [25, 28, DIFF and highest ACC​ (see line 4 in Table 1). Thus, the
29], and they are denoted by L-PM and G-PM, respectively. best solution provided by G-PM is the closest to the DM’s
Two criteria are used to evaluate the effectiveness of most satisfactory solution; however, G-PM is most capable
the different preference models. The first criterion is that of capturing the DM’s preferences. This is because the DM’s
adopted in the literature [25, 28, 29], i.e., the difference/dis- preference is formulated by a specific polynomial function,
tance between the DM’s truly preferred solution determined which can be easily adapted by G-PM, a general polynomial
by the preference function and the best solution produced function. The DIFF and ACC​values produced by RFFS-DT
by the preference model, denoted by DIFF, and the other (see line 2 in Table 1) show that our proposed dynamic fea-
is the accuracy of the preference model, denoted by ACC​. ture analysis-based decision tree preference model can also
The accuracy of the preference model is defined as the per- approximate the DM’s polynomial function-based prefer-
centage of currently nondominated solutions with rankings ence. L-PM (see line 3 in Table 1) shows the worst perfor-
(given by the DM’s preference feedback) that are correctly mance and illustrates that if the function used for modeling
predicted by the preference model. Accuracy represents the
overall reflection of the preference model’s capability to cap-
Table 1  Comparative results under a polynomial function-based pref-
ture the DM’s preference feedback.
erence
The values of the parameters described in Sect. 4 are
set as follows: s = 100, max_t = 100, max_it = 20, gen = 5 Preference models DIFF ACC​
and st = 5. The MV model is a two-objective optimization RFFS-DT 24.65E−09 0.925
problem. We find that a normal MOEA/D with a popula- L-PM 573.05E−09 0.4416
tion size s = 100 and a maximum number of generations G-PM 1.85E−09 0.9845
max_t = 100 can obtain a set of nondominated solutions that
is very close to the true Pareto front. Since our aim is not
to boost MOEA/D performance by fine tuning its param-
Table 2  Comparative results under an exponential function-based
eters but rather to show the improvements achieved by our
preference
novel RFFS-DT for IMCDM, as in [28, 29], we give only the
default values of the two parameters that are usually associ- Preference models DIFF ACC​
ated with the evolutionary algorithm and focus on the three RFFS-DT 0.55E−09 0.936
parameters that are specifically introduced to IMCDM and L-PM 227.12E−09 0.6735
directly impact its performance. First, we choose the default G-PM 5.87E−09 0.84
value for st, which is the number of solutions presented to
the DM for pairwise comparison. Of course, if additional
solutions are offered for feedback, a more appropriate pref-
Table 3  Comparative results under a logarithmic function-based pref-
erence model can be derived; however, this choice will sig-
erence
nificantly increase the DM’s cognitive load. To maintain a
reasonable per-interaction burden on the DM, st is fixed to Preference models DIFF ACC​
5 in the IMCDM [28, 29] ([29] used the same evolution- RFFS-DT 1.79E−07 0.96
ary algorithm MOEA/D as our IMCDM). As our IMCDM L-PM 257.76E−07 0.5468
adopted the same form of preference feedback as these G-PM 12.57E−07 0.82

13
International Journal of Machine Learning and Cybernetics

preference has a different form from the one used for the Furthermore, in contrast to the results shown in Table 1,
DM’s preference, it is impossible to capture the DM’s true the results in Table 4 show that the G-PM’s performance
preference and unlikely to return a satisfactory solution to deteriorates to a greater degree than that of RFFS-DT, thus
the DM. In other cases, the DM’s preference is shaped by confirming that the functional structure has more influence
complex functions such as exponential and logarithmic func- than the functional parameter in terms of modeling the DM’s
tions rather than a polynomial function; the results of the preference.
comparison are provided in Tables 2 and 3 show that our
proposed RFFS-DT exhibits the best capability to adapt to (2) DM with variable preferences
the DM’s preference. Any polynomial function-based pref-
erence model, whether it is L-PM or G-PM, is not suitable In this case, the DM’s preference feedback is given based
for modeling the DM’s preference with respect to the DIFF on functions (8), (9) and (10) for interaction intervals [1,
and ACC​values. 10], [11, 20] and [21, 40], respectively. We compare only
Tables 4, 5, and 6 present the results of the comparison the performance of RFFS-DT and G-PM since L-PM has
under the DM’s preferences formulated by functions (11), been proven to be the worst preference model in the simple
(12) and (13), respectively. These preference functions case of a DM with a stable preference. Table 7 shows the
include an additional impact factor term. The performance results of the comparison. It is seen from the DIFF and ACC​
of L-PM and G-PM deteriorate considerably, while the values that RFFS-DT can better capture the DM’s variable
RFFS-DT maintains nearly the same level of performance as preferences than G-PM.
before. As mentioned earlier, both L-PM and G-PM have a To gain insight into how the two preference models adapt
predetermined form (composed of only objective functions) to variable preferences during the decision-making process,
and cannot embed new terms; therefore, if the change in we present the ACC​values for each interaction (see Fig. 3).
the DM’s preference is caused by additional impact factors, In the interaction interval [1, 10], the G-PM is more capa-
they cannot be adjusted to fit the DM’s true preference only ble of being adapted to the DM’s preference than RFFS-DT
by updating the coefficient of each term. However, since since the DM’s preference is restricted to a polynomial func-
the RFFS-DT is built based on dynamic feature analysis, its tion, which is consistent with the previous result shown in
structure can adapt to any change in the DM’s preferences. Table 1. In the interaction intervals [11, 20] and [21, 40],
the RFFS-DT exhibits remarkable superiority in modeling
the DM’s preference using a nonpolynomial function, which
Table 4  Comparative results under a polynomial function-based pref- coincides with the results shown in Tables 2 and 3. Fur-
erence with an additional impact factor
thermore, when the DM’s preference changes at interactions
Preference models DIFF ACC​ no. 11 and no. 21, the ACC​ values drop sharply to nearly
zero but increase in the following interactions. Obviously,
RFFS-DT 25.61E−09 0.9162
RFFS-DT could recover to a much higher degree than G-PM
L-PM 946.13E−09 0.3205
between each pair of adjacent interactions.
G-PM 30.25E−09 0.7131
The results of a comparison between RFFS-DT and
G-PM are illustrated in Table 8 and Fig. 4 when the DM’s
preference feedback is given based on functions (11), (12)
Table 5  Comparative results under an exponential function-based and (13) for the interaction intervals [1, 10], [11, 20] and
preference with an additional impact factor
[21, 40], respectively. These preference functions are affili-
Preference models DIFF ACC​ ated with the terms of additional impact factors. IN contrast
to the results shown in Table 7, the results in Table 8 show
RFFS-DT 0.57E−09 0.929
that the performance of RFFS-DT slightly decreases but that
L-PM 602.08E−09 0.4132
G-PM is considerably worse, which demonstrates the sig-
G-PM 13.62E−09 0.671
nificant effectiveness of dynamic feature analysis in building
the RFFS-DT. Furthermore, as shown in Fig. 3, the results of
the comparison shown in Fig. 4 reveal that, under dynamic
Table 6  Comparative results under a logarithmic function-based pref- feature analysis, RFFS-DT can resume to a very high level of
erence with an additional impact factor
Preference models DIFF ACC​ Table 7  Comparative results under variable preferences
Preference models DIFF ACC​
RFFS-DT 1.66E−07 0.949
L-PM 559.98E−07 0.4014 RFFS-DT 0.00024025 0.89
G-PM 30.03E−07 0.603 G-PM 0.00036100 0.5468

13
International Journal of Machine Learning and Cybernetics

Fig. 3  Accuracy changes in the different stages of interactions

Table 8  Comparative results under variable preferences with addi- 5.3 Study of the effect of critical parameters
tional impact factors
Preference models DIFF ACC​ The effects of the two critical parameters, max_it and gen,
on our IMCDM is studied using the default values of the
RFFS-DT 0.00029307 0.8766
other three parameters, i.e., s, max_t and st. These three
G-PM 0.00078174 0.4380
default values were previously determined empirically as
well as analytically. Since max_t can be defined as a product
of max_it and gen, the values of parameters max_it and gen
accuracy in the same interaction period after a sharp change can be derived. Currently, we can study only the effect of
in the DM’s preference. Due to its predetermined functional either parameter, e.g., gen. In our parametric study, we select
structure, G-PM can resume to a level of accuracy that is the values of gen from the range [1, 10]. For each value of
much lower than that of RFFS-DT. gen, we test RFFS-DT on each of the preference functions
(8)–(13). For each value of gen and each preference function,

Fig. 4  Accuracy changes in the different stages of interaction with additional impact factors

13
International Journal of Machine Learning and Cybernetics

our IMCDM is executed in several iterations, and then the values for the parameters in IMCDM [23, 25]. For exam-
performance indices DIFF and ACC​are averaged. Figures 5 ple, the above figures show the best performance occurs
and 6 give the DIFF and ACC​ values changing with the when gen = 1; however, the DM must provide his or her
values of gen. In each figure, (a–f) correspond to preference preference information after each iteration, and the total
functions (8)–(13), respectively. number of max_t interactions is typically beyond the scope
From these figures, we can see that, in general, the perfor- of a DM’s cognitive capability [26, 56]. Therefore, an
mance of IMCDM decreases as the value of gen increases (a appropriate value for gen involves a compromise between
large value of gen leads to a small number of interactions). the number of interactions and the performance measures.
The rationale is that when the DM interacts more frequently, Figures 5 and 6 show that, in most cases, both DIFF and
the preference model RFFS-DT will be updated more often, ACC​ decrease mildly until gen = 5, i.e., the maximum
which means that the DM’s preference information can be number of interactions, max_it = 20. Hence, these two
captured by the preference model in a timely and precise values are selected as the default values in the parameter
manner, and as a result, a more satisfactory solution can be initialization for our comparative experiments.
found by using the preference model than by using other
models. Nevertheless, if the DM interacts less frequently, the
preference model will capture little preference information, 6 Conclusions and future directions
making it impossible for the search to focus on the DM’s
preferred solution. For interactive portfolio optimization, this paper pro-
In addition to performance measures, cognitive load poses a self-adaptive preference model based on dynamic
criteria must be considered when selecting the appropriate feature analysis, RFFS-DT, to capture the DM’s complex

Fig. 5  DIFF values change with the values of gen

13
International Journal of Machine Learning and Cybernetics

Fig. 6  ACC​values change with the values of gen

preferences during the decision-making process. We first available. However, in a real application such as portfolio
define three categories of impact factors that may influence optimization, it is unrealistic to assume a certain structure
the DM’s preferences and designed a random-forest-based form and impact factors of the DM’s preference in advance.
feature selection method. After each interaction during the Thus, the self-adaptiveness of our proposed RFFS-DT pro-
decision process, RFFS is applied to the DM’s preference vides an effective solution for such complicated scenarios.
feedback to obtain a set of solutions for discerning the DM’s In terms of machine learning, IMCDM is a process of
true preference impact factors. Then, RFFS-DT is built and target concept (i.e., DM’s true preference) learning from the
updated in terms of the selected features (i.e., impact fac- DM’s preference feedback. Similar to the IMCDMs in the
tors). In contrast to the current function-based preference literature, our IMCDM adopts a passive learning strategy;
models used in the literature, RFFS-DT is based on the very i.e., at each interaction, a certain number of solutions are
general assumption that the DM’s preference structure is randomly selected from the current set of nondominated
unknown a priori and that his or her preference impact fac- solutions for class labeling and preference learning. The
tors may change unpredictably during the decision-making randomly selected sample by passive learning is focused on
process. A comparison of the experimental results demon- diversity but is short of informativeness. Recent studies have
strates the effectiveness of our proposed RFFS-DT, espe- shown that by incorporating diversity and informativeness
cially when no information about the DM’s preferences is into the sample selection, the number of examples needed

13
International Journal of Machine Learning and Cybernetics

for learning a target concept can be greatly decreased [58, An appropriate compromise between the generalizability
59]. In future research on our IMCDM, this sampling strat- and uncertainty [62, 63] of a preference model will make
egy-based active learning may provide an innovative method the model robust to the DM’s variable preferences without
to reduce the number of solutions presented to the DM for any prior knowledge in the decision process and hence help
preference feedback at each interaction. reduce the number of interactions in our IMCDM.
The decision tree-based preference model exhibits more
flexibility in modeling DM’s complicated preferences than Acknowledgments The authors thank the referee for his or her com-
ments on the manuscript. This work was supported in part by the
the function-based preference models. However, due to the NSFS under Grant No. ZR2016FM27and by the NSFC under Grant
intrinsic overfitting, it is unlikely to give full consideration No. 61702139.
of the DM’s preference uncertainty in the decision-mak-
ing process. Ambiguity-based active learning [60, 61] is
an alternative to building a more robust preference model
for our IMCDM. Because the preference model is used to Appendix
predict the class label for the next generated nondominated
solutions, it should retain considerable generalizability while See Table 9.
taking into consideration the DM’s preference uncertainty.

Table 9  The 18 selected stocks Stock category and code Company Company in short Expected return
and their returns stock code

Agriculture, forestry, husbandry, fishery (A) 000998 Longping HiTech 0.017050


Mining (B) 600028 Sinopec 0.015300
Manufacturing (C) 600519 Guizhou Moutai 0.065458
Electricity, heating, gas, water, supply (D) 600900 Yangtze Power 0.021958
architecture (E) 002310 Eastern garden 0.032633
Wholesale, retailer (F) 002251 BBK 0.011300
Transportation, storage, post (G) 601111 Air China 0.049533
Accommodation, catering (H) 600754 Jinjiang stake 0.010600
Information (I) 002230 iFlytek 0.076600
Finance (J) 600036 CMBC 0.046683
Estate (K) 000002 Shenzhen Vanke 0.040292
Lease, business service (L) 601888 CITS 0.063708
Science, technique (M) 002178 Yanhua Intel 0.038308
Irrigation, environment, infrastructure (N) 000069 OCT A 0.020350
Education (P) 603377 Oriental Fashion 0.002617
Sanitation, society (Q) 300015 Aier Eye 0.038658
Culture, PE, entertainment (R) 002624 Excel Glob 0.011650
Composite (S) 600455 Broadcom shares 0.057975

13
International Journal of Machine Learning and Cybernetics

References 22. Miettinen K, Mäkelä MM, Kaario K (2006) Experiments with


classification-based scalarizing functions in interactive multi-
objective optimization. Eur J Oper Res 175(2):931–947
1. Markowitz H (1952) Portfolio selection. J Financ 7(1):77–91
23. Luque M, Miettinen K, Eskelinen P, Ruiz F (2009) Incorporating
2. Markowitz H (1959) Portfolio selection: efficient diversification
preference information in interactive reference point methods for
of investments. John Wiley and Sons, New York
multi-objective optimization. Omega 37(2):450–462
3. Deb K, Pratap A, Agarwal S, Meyarivan T (2002) A fast and elit-
24. Vallerio M, Hufkens J, Van IJ, Logist F (2015) An interactive
ist multi-objective genetic algorithm: NSGA-II. IEEE Trans Evol
decision-support system for multi-objective optimization of
Comput 6(2):182–197
nonlinear dynamic processes with uncertainty. Expert Syst Appl
4. Zhang Q, Li H (2007) MOEA/D: a multi-objective evolution-
42(21):7710–7731
ary algorithm based on decomposition. IEEE Trans Evol Comput
25. Battiti R, Passerini A (2010) Brain-computer evolutionary multi-
11(6):712–731
objective optimization: a genetic algorithm adapting to the deci-
5. Xin B, Chen L, Chen J, Ishibuchi H, Hirota K, Liu B (2018) Inter-
sion maker. IEEE Trans Evol Comput 14(5):671–687
active multiobjective optimization: a review of the state-of-the-art.
26. Larichev O (1992) Cognitive validity in design of decision aiding
IEEE Access 6:41256–41279
techniques. J Multi-Criteria Decis Anal 1(3):127–138
6. Bonissone PP, Subbu R, Lizzi J (2009) Multi-criteria decision
27. Pascoletti A, Serafini P (1984) Scalarizing vector optimization
making: a framework for research and applications. IEEE Comput
problems. J Optim Theory Appl 42(4):499–524
Intell Mag 4(3):48–61
28. Deb K, Sinha A, Korhonen PJ, Wallenius J (2010) An interac-
7. Ozbey O, Karwan MH (2014) An interactive approach for mul-
tive evolutionary multi-objective optimization method based on
ticriteria decision making using a Tchebycheff Utility function
progressively approximated value functions. IEEE Trans Evol
approximation. J Multi-Criteria Decis Anal 21(3–4):153–172
Comput 14(5):723–739
8. Ruiz AB, Saborido R, Luque M (2015) A preference-based evo-
29. Mukhlisullina D, Passerini A, Battiti R (2013) Learning to diver-
lutionary algorithm for multi-objective optimization: the weight-
sify in complex interactive multi-objective optimization. Proceed-
ing achievement scalarizing function genetic algorithm. J Global
ings of the 10th Metaheuristics International Conference, Singa-
Optim 62(1):101–129
pore Management University pp. 230–239
9. Li L, Wang Y, Trautmann H, Jing N, Emmerich M (2018) Multi-
30. Sun M, Stam A, Steuer RE (2000) Interactive multiple objective
objective evolutionary algorithms based on target region prefer-
programming using tchebycheff programs and artificial neural
ences. Swarm Evol Comput 40:196–215
networks. Comput Oper Res 27(7):601–620
10. Li L, Yevseyeva I, Basto-Fernandes V, Trautmann H, Jing N
31. Sun M, Stam A, Steuer RE (1996) Solving multiple objective pro-
(2017) Building and using an ontology of preference-based mul-
gramming problems using feed-forward artificial neural networks:
tiobjective evolutionary algorithms. Int Conf Evol Multi-Criterion
the interactive FFANN procedure. Manag Sci 42(6):835–849
Optim 10173:406–421
32. Huang HZ, Tian Z, Zuo MJ (2005) Intelligent interactive multi-
11. Köksalan M, Wallenius J, Zionts S (2013) An early history of
objective optimization method and its application to reliability
multiple criteria decision making. J Multi-Criteria Decis Anal
optimization. IIE Trans 37(11):983–993
20(1–2):87–94
33. Zhang H, Llorca J, Davis CC et al (2012) Nature-inspired self-
12. Phelps SP, Koksalan M (2003) An interactive evolutionary
organization, control, and optimization in heterogeneous wireless
metaheuristic for multi-objective combinatorial optimization.
networks. IEEE Trans Mob Comput 11(7):1207–1222
Manag Sci 49(12):1726–1738
34. Zhang H, Cao X, Ho JKL et al (2017) Object-level video adver-
13. Branke J, Corrente S, Greco S, Słowi´nski R, Zielniewicz P
tising: an optimization framework. IEEE Trans Industr Inf
(2016) Using choquet integral as preference model in interac-
13(2):520–531
tive evolutionary multiobjective optimization. Eur J Oper Res
35. Qamar N, Akhtar N, Younas I (2018) Comparative analysis of
250(3):884–901
evolutionary algorithms for multi-objective travelling salesman
14. Pedro LR, Takahashi RHC (2014) INSPM: an interactive evolu-
problem. Int J Adv Comput Sci Appl 9(2):371–379
tionary multi-objective algorithm with preference model. Inf Sci
36. Masmoudi M, Abdelaziz FB (2018) Portfolio selection problem: a
268(1):202–219
review of deterministic and stochastic multiple objective program-
15. Greco S, Ehrgott M, Figueira JR (2016) Multiple criteria decision
ming models. Ann Oper Res 267(1–2):335–352
analysis state of the art surveys. Springle, New York
37. Källblad S (2017) Risk-and ambiguity-averse portfolio opti-
16. Miettinen K, Mäkelä MM (2002) On scalarizing functions in mul-
mization with quasicon-cave utility functionals. Financ Stoch
tiobjective optimization. OR Spectr 24(2):193–213
21(2):397–425
17. Deb K, Sundar J (2006) Reference point based multi-objective
38. Lwin KT (2015) Evolutionary approaches for portfolio optimiza-
optimization using evolutionary algorithms. Proceedings of the
tion. Dissertation, University of Nottingham
8th ACM annual conference on Genetic and evolutionary com-
39. Gutjahr WJ, Pichler A (2016) Stochastic multi-objective opti-
putation pp. 635–642
mization: a survey on non-scalarizing methods. Ann Oper Res
18. Klamroth K, Miettinen K (2008) Integrating approximation and
236(2):475–499
interactive decision making in multicriteria optimization. Oper
40. Brucker P, Knust S (2012) Resource-constrained project schedul-
Res 56(1):222–234
ing in: complex scheduling. GOR-Publications
19. Wang R, Purshouse RC, Fleming PJ (2013) Preference-inspired
41. Miettinen K, Ruiz F, Wierzbicki AP (2008) Introduction to multi-
coevolutionary algorithms for many-objective optimization. IEEE
objective optimization: interactive approaches. In: Multiobjective
Trans Evol Comput 17(4):474–494
Optimization, pp 27–57
20. Goulart F, Campelo F (2016) Preference-guided evolutionary
42. Breiman L (2001) Random forests. Mach Learn 45(1):5–32
algorithms for many-objective optimization. Inf Sci 329:236–255
43. Biau G, Devroye L, Lugosi G (2008) Consistency of ran-
21. Miettinen K, Mäkelä MM (2006) Synchronous approach
dom forests and other averaging classifiers. J Mach Learn Res
in interactive multi-objective optimization. Eur J Oper Res
9(1):2015–2033
170(3):909–922
44. Mentch L, Hooker G (2014) Quantifying uncertainty in random
forests via confidence intervals and hypothesis tests. J Mach Learn
Res 17(1):841–881

13
International Journal of Machine Learning and Cybernetics

45. Verikas A, Gelzinis A, Bacauskiene M (2011) Mining data with 56. Goulart F, Campelo F (2016) Preference-guided evolutionary
random forests: a survey and results of new tests. Pattern Recogn algorithms for many-objective optimization. Inf Sci 329:236–255
44(2):330–349 57. Saaty TL, Vargas LG (2017) Models, methods, concepts &
46. Marill T, Green D (1963) On the effectiveness of receptors in applications of the analytic hierarchy process. International
recognition systems. IEEE Trans Inf Theory 9(1):11–17 7(2):159–172
47. Grabczewski K (2014) Meta-learning in decision tree induction. 58. Wang R, Wang XZ, Kwong S, Xu C (2017) Incorporating diver-
Springer International Publishing, Berlin sity and informativeness in multiple-instance active learning.
48. Safavin SR, Landgrebe D (1991) A survey of decision tree clas- IEEE Trans Fuzzy Syst 25(6):1460–1475
sifier methodology. IEEE Trans Syst Man Cybern 21(3):660–674 59. Du B, Wang ZM, Zhang LF, Zhang LP, Liu W, Shen JL, Tao
49. Zhang H, Singer BH (2010) Recursive partitioning and applica- DC (2015) Exploring representativeness and informativeness for
tions. Springer, New York active learning. IEEE Trans Cybern 47(1):1–13
50. Breiman L, Friedman J, Olshen R, Stone C (1984) Classification 60. Wang R, Chow CY, Kwong S (2016) Ambiguity based multiclass
and regression trees. Chapman & Hall, New York active learning. IEEE Trans Fuzzy Syst 24(1):242–248
51. Wang R, Kwong S, Wang XZ, Jiang QS (2015) Segment based 61. Samat A, Gamba P, Liu SC (2018) Fuzzy multiclass active
decision tree induction with continuous valued attributes. IEEE learning for hyperspectral image classification. IET Image Proc
Trans Cybern 45(7):1262–1275 12(7):1–7
52. Bosman PAN, Thierens D (2003) The balance between proximity 62. Wang XZ, Wang R, Xu C (2018) Discovering the relationship
and diversity in multi-objective evolutionary algorithms. IEEE between generalization and uncertainty by incorporating complex-
Trans Evol Comput 7(2):174–188 ity of classification. IEEE Trans Cybern 48(2):703–715
53. Fowler JW, Gel ES, Koksalan MM, Korhonen P, Marquis JL, 63. Wang XZ, Xing HJ, Li Y, Hua Q, Dong CR, Pedrycz W (2015) A
Wallenius J (2010) Interactive evolutionary multi-objective opti- study on relationship between generalization abilities and fuzzi-
mization for quasi-concave preference functions. Eur J Oper Res ness of base classifiers in ensemble learning. IEEE Trans Fuzzy
206(2):417–425 Syst 23(5):1638–1854
54. Sinha A, Korhonen P, Wallenius J, Deb K (2010) An interactive
evolutionary multi-objective optimization method based on poly- Publisher’s Note Springer Nature remains neutral with regard to
hedral cones. Learn Intell Optim 6073(3):318–332 jurisdictional claims in published maps and institutional affiliations.
55. Koksalan M, Karahan I (2010) An interactive territory defin-
ing evolutionary algorithm: iTDEA. IEEE Trans Evol Comput
14(5):702–722

13

You might also like