Conference PaperPDF Available

Fisher Score-Based Feature Selection for Ordinal Classification: A Social Survey on Subjective Well-Being


Abstract and Figures

This paper approaches the problem of feature selection in the context of ordinal classification problems. To do so, an ordinal version of the Fisher score is proposed. We test this new strategy considering data from an European social survey concerning subjective well-being, in order to understand and identify the most important variables for a person’s happiness, which is represented using ordered categories. The input variables have been chosen according to previous research, and these have been categorised in the following groups: demographics, daily activities, social well-being, health and habits, community well-being and personality/opinion. The proposed strategy shows promising results and performs significantly better than its nominal counterpart, therefore validating the need of developing specific ordinal feature selection methods. Furthermore, the results of this paper can shed some light on the human psyche by analysing the most and less frequently selected variables.
Content may be subject to copyright.
Fisher Score-Based Feature Selection
for Ordinal Classification: A Social Survey
on Subjective Well-being?
M. P´erez-Ortiz1, M. Torres-Jim´enez1, P.A. Guti´errez2, J. S´anchez-Monedero1,
and C. Herv´as-Mart´ınez2
1Universidad Loyola Andaluc´ıa, Dept. of Quantitative Methods, C´ordoba, Spain,,
2University of C´ordoba, Dept. of Computer Science and Numerical Analysis,
ordoba, Spain
Abstract. This paper approaches the problem of feature selection in the
context of ordinal classification problems. To do so, an ordinal version
of the Fisher score is proposed. We test this new strategy considering
data from an European social survey concerning subjective well-being,
in order to understand and identify the most important variables for a
person’s happiness, which is represented using ordered categories. The
input variables have been chosen according to previous research, and
these have been categorised in the following groups: demographics, daily
activities, social well-being, health and habits, community well-being and
personality/opinion. The proposed strategy shows promising results and
performs significantly better than its nominal counterpart, therefore val-
idating the need of developing specific ordinal feature selection methods.
Furthermore, the results of this paper can shed some light on the human
psyche by analysing the most and less frequently selected variables.
1 Introduction
The nature of well-being is a topic that has exercised the minds of moral philoso-
phers for centuries [1]. Recently, research on happiness has gained importance,
not only in the psychology area, but also in other fields like economics [2]. A
number of nations have begun to develop measures of subjective well-being [3]
to complement traditional measures of national well-being, such as the Gross
Domestic Product. Well-being research is usually clustered into two camps [1],
focusing either on subjective well-being or psychological well-being. On the one
hand, subjective well-being is understood as having an emotional component of
the balance between positive and negative affect and a cognitive component of
judgements about life satisfaction. On the other hand, psychological well-being
?This work has been subsidised by the TIN2014-54583-C2-1-R project of the Spanish
Ministerial Commission of Science and Technology (MICYT), FEDER funds and
the P11-TIC-7508 project of the “Junta de Andaluc´ıa” (Spain).
has been defined as “engagement with existential challenges of life” [4]. Given
the diversity of perspectives on the definition of subjective and psychological
well-being, it is not surprising that different measurements have been considered
in each case. In empirical research, a number of studies suggests that subjective
and psychological well-being are two related, but distinct, constructs [4].
Nowadays, machine learning represents one of the most actively researched
technical fields, mainly because of its applicability to very different domains.
In this sense, machine learning, which lies at the intersection of computer sci-
ence and statistics and is at the core of artificial intelligence and data science,
addresses the question of how to build computer-based systems that improve
automatically through experience. Given the lack of empirical agreement on
the structure of well-being and the use of non-validated measures in previous
studies, the current study examines these issues using machine learning tech-
niques and data from the European Social Survey (ESS)1. The ESS includes a
large sample of European inhabitants and validated well-being measures. It is
an academically driven cross-national survey and has been conducted every two
years across Europe. The survey measures the attitudes, beliefs and behaviour
patterns of diverse populations in more than thirty nations, involving strict ran-
dom probability sampling, a minimum target response rate of 70% and rigorous
translation protocols. The interviews include questions on a variety of core top-
ics: social trust, political interest and participation, socio-political orientations,
social exclusion, national, ethnic and religious allegiances, health and social de-
terminants, immigration, human values, demographics and socioeconomics.
This paper presents a new strategy to perform feature selection in ordinal
classification [5]. Ordinal classification comprises those classification problems
where the variable to predict follows a natural order (e.g. in Likert scales, as
the case considered here). More specifically, we develop a novel feature selection
methodology based on the well-known Fisher score [6]. The proposed technique
promotes features that maintain the order among the classes and is used in this
paper to analyse which are the factors influencing subjective well-being to a
larger extent.
The remainder of the paper is structured as follows: Section II presents the
data considered; Section III presents some previous notions and the proposed
strategy for feature selection; Section IV analyses and presents the results ob-
tained; and finally, Section V outlines some conclusions and final remarks.
2 Social survey on subjective well-being
The survey data conducted in 2014 from 15 European Union countries have been
selected according to the availability of information (all persons aged 15 and
over, resident within private households, regardless of their nationality, citizen-
ship, language or legal status, in the following participating countries: Austria,
Belgium, Czech Republic, Denmark, Estonia, Finland, France, Germany, Ire-
land, Netherlands, Slovenia, Sweden and Switzerland). Different variables have
been selected to predict the level of well-being of European citizens, considering
the components that influences happiness according to previous research [7, 8].
These variables have been classified into six different groups:
– Demographics: different factors including country, age, gender, education,
familiar composition, financial matters, etc.
– Daily activities: this group considers different variables that indicate how
people spend their time (e.g. number of working hours, main activity, em-
ployment contract, number of hours watching TV, etc.).
– Social well-being: these variables are related to the social environment of
the person (e.g. how often they take part in social activities, the number
of people they are living with, how many people they can discuss personal
matters with, etc.).
Health and habits: including how often they practice sports, how often they
eat vegetables, subjective general health, how often they drink alcohol, smok-
ing behaviour, quality of sleep, and others.
Community well-being: related to the sense of engagement they have with the
area they live. It includes politic and environmental aspects (e.g. whether
they feel close to their country, how interested they are in politics, how
satisfied they are with economy/health services/education/government in
their country, placement on left-right scale, etc.).
Personality/opinion: how religious they are, whether it is important to be
rich/free/humble/adventurous, whether they would allow immigrants from
poorer countries to settle in their country, whether most people can be
trusted or not, etc.
The study comprise a set of 56 different variables: a large number of them (38)
represent Likert scales (i.e. ordinal) and are codified using a numeric scale, 7 are
numeric, 6 of them are binary, and finally, there are 5 nominal variables, which
are transformed to binary ones, resulting then in a total set of 91 variables. The
total number of cases is 28,137, excluding those with responses “don’t know”,
“no answer” or “refusal” in the dependent variable, which is how happy they
are in a Likert scale (from 0 to 10). Three different datasets are considered,
using different number of levels for the subjective well-being: all the 11 possible
answers (referred to as SW-11c), 5 classes (SW-5c, where C1={0,1,2},C2=
{3,4},C3={5,6},C4={7,8}and C5={9,10}) and 3 classes (SW-3c, where
C1={0,1,2,3},C2={4,5,6,7}and C3={8,9,10}).
Values such as “don’t know”, “no answer” or “refusal” in the independent
variables have been considered as missing values and have been imputed using
the Event Covering algorithm [9], as suggested in [10] for approximate models
such as neural networks, support vector machines and other statistical methods.
3 Methodology
This section describes some previous notions (the paradigm of ordinal classifica-
tion and the Fisher score) and proposes an ordinal feature selection method.
3.1 Previous notions
Ordinal classification and ordinal feature selection The classification of
patterns into naturally ordered labels is referred to as ordinal regression or or-
dinal classification. This learning paradigm, although still mostly unexplored, is
spreading rapidly and receiving a lot of attention from the pattern recognition
and machine learning communities [5,11], given its applicability to real world
problems. This paradigm shares properties of classification and regression. In
contrast to multinomial classification, there exists some ordering among the el-
ements of Y(the labelling space) and both standard classifiers and the zero-one
loss function do not capture and reflect this ordering. Concerning regression, Y
is a non-metric space, so the distances between categories are not known.
As an explanatory example, consider the case of financial trading where an
agent predicts not only whether to buy an asset, but also the investment. The
different situations could be categorised as {“no investment”, “little investment”,
“big investment”, “huge investment”}. A natural order among the classes exists
in this case, and a necessity of penalising differently the misclassification errors
(it should not be considered equal misclassifying a “no investment” instance with
a “huge investment” one than misclassifying it with “little investment”).
The goal in ordinal classification is to assign an input vector xto one of K
discrete classes Ck, k ∈ {1, . . . , K}, where there exists a given ordering between
the labels, C1≺ C2≺ · · · ≺ CK,denoting this order information. Hence, the
objective is to find a prediction rule C:X → Y by using an i.i.d. training sample
X={xi, yi}N
i=1, where Nis the number of training patterns, xi X ,yi∈ Y,
X Rdis the d-dimensional input space and Y={C1,C2,...,CK}is the label
space. For convenience, denote by Xito the set of patterns belonging to Ci.
Despite the novelty of ordinal classification, there is some research concern-
ing new prediction strategies for these problems. However, there are aspects of
ordinal classification that are receiving less attention. This is the case of feature
selection methods in ordinal classification, where the number of approaches is
still low [12, 13]. Like some of the strategies in the feature selection literature,
these techniques rely on a discretisation of the input space. In this paper, we
devise a new strategy for ordinal feature selection that is free of this requirement.
Fisher score for feature selection The problem of supervised feature se-
lection is now described. Given a dataset {xi, yi}N
i=1, we aim to find a feature
subset of size m(where m<d) that contains the most informative features.
The Fisher score for feature selection [6] was proposed as an heuristic strategy
for computing an independent score for each feature using the well-known notion
of the Fisher ratio. Let µi
kand σi
kbe the mean and standard deviation of the
k-th class and i-th feature (and µiand σithe mean and standard deviation of
the whole dataset for the i-th feature). The Fisher score for the i-th feature (xi)
can be computed as:
F(xi) = PK
k=1 Nk(µi
k=1 Nk(σi
where Nkis the number of patterns of class Ck. Since this score is computed
independently, the features selected may represent a suboptimal set. Further-
more, this heuristic may fail to select redundant features or those with a high
aggregated discriminative power. This technique is named as Nominal Feature
Selection (NFS) in the experimental results.
3.2 Proposed feature selection strategy
In this paper, we reformulate the nominal Fisher score Fto deal with ordinal
data (named it as FO). In this regard, we include a weighting term that intro-
duces a higher cost for distant classes. This cost will force the feature selection
method to focus more on those features that help to discriminate classes that are
far in the ordinal scale (in order to avoid the above-mentioned misclassification
errors). The formulation proposed is the maximise the following score:
FO(xi) = PK
k=1 PK
j=1 |kj|(µi
(K1) PK
where |kj|is the cost of misclassifying a pattern from class Ckin class Cj.
Apart from the fact that more distant classes in the ordinal scale should
present a higher distance between them, there is another ordinal requirement
which can be introduced in the feature selection stage. As said before, the la-
belling space is non-metric, therefore we can not introduce a distance relation
among the different classes. However, from the ordinal classification definition
it can be stated that d(Ck,Cj)< d(Ck,Ch),∀ {k, j, h, |k6=j6=h(k < j < h
k > j > h)}. These ordinal requirements can be introduced by the score:
OR(xi) = PK2
k=1 PK1
j=k+1 PK
j=2 (Kj),(3)
being J·Ka Boolean test which is 1 if the inner condition is true, and 0 otherwise.
This ORscore measures the number of ordinal requirements fulfilled for a specific
feature i. To include both terms in the feature selection process, we compute a
weighted mean of both scores in the following manner:
FOR(xi) = α·FO(xi) + (1 α)OR(xi),(4)
where α(0,1).
Up to now, we have defined the distance between the classes Czand Cjas
j)2(note that this formulation presents problems with non-
linear, multimodal or non-normal data). Alternative metrics have been proposed
for measuring the distance between classes [14]. In this way, we consider other
notions of distance between sets of points, such as the mean distance:
m(Cz,Cj) = 1
where the idea is to compute the mean distance between each pair of patterns
of different classes. Another alternative is the sum of minimum distances:
md(Cz,Cj) = 1
h,Cj) + X
where m(xi
h,Cj) = minxi
v) and drepresents the Euclidean distance.
Finally, the Hausdorff distance is defined as:
h(Cz,Cj) = max{max
In the experiments of this paper, we will test these three alternative approaches
for computing inter-class distances.
4 Experimental results
This section exposes the experiments considered in this paper and analyses the
results obtained. Regarding the experimental setup, a stratified holdout tech-
nique was applied to randomly divide the datasets 30 times, using 5% of the
patterns for training (approximately 1400 patterns) and the remaining 95% for
testing. The partitions were the same for all methods and one model was ob-
tained and evaluated (in the test set) for each split. Finally, the results are
computed as the mean and standard deviation of the measures over the 30 test
sets. The classification method chosen is the reformulation of Kernel Discrimi-
nant Analysis for Ordinal Regression (KDLOR), because of its relation to the
Fisher score [15].
4.1 Methodologies tested
Different methods are compared:
No feature selection. These results are obtained with the KDLOR classifica-
tion method using the whole set of features.
Nominal Feature Selection (NFS). In this case, we use the nominal version
of the Fisher score.
Ordinal Feature Selection (OFS) with dsas distance metric.
Ordinal Feature Selection using the mean distance (OFSMean). In this case,
dmis used as distance metric.
Ordinal Feature Selection using the min distance (OFSMin). In this case,
dmd is used as distance metric.
Ordinal Feature Selection using the Hausdorff distance (OFSHausdorff ). In
this case, dhis used as distance metric.
The value αfor all the ordinal feature selection methods is fixed to α= 0.5.
For all feature selection methods, features are ranked according to the corre-
sponding score, and then a percentage of the best features is retained (where
the percentages tested are 10%,20%,...,90%). Concerning the parameters se-
lected for the classification method, a Gaussian kernel is used for KDLOR,
K(x,y) = exp kxyk2
σ2, where σis the kernel width which has been cross-
validated using a 5-fold nested procedure over the training set and the range
4.2 Evaluation metrics
Several measures can be considered for evaluating ordinal classifiers, such as
the Mean Absolute Error (MAE) or the accuracy or Acc [5]. In this work, two
metrics have been used:
Acc is the ratio of correctly classified patterns:
Acc =1
where yiis the desired output for pattern iand y
iis the prediction.
Acc and MAE may not be the best option when measuring performance
in the presence of class imbalances [16]. The average mean absolute error
(AMAE) is the mean of MAE classification errors throughout the classes,
where MAE is the average absolute deviation of the predicted class from
the true class. Let MAEkbe the M AE for a given k-th class:
|O(yi)− O(y
where O(Ck) = k, 1 kK, i.e. O(yi) is the order of class label yi. Then,
the AMAE measure can be defined in the following way:
MAE values range from 0 to K1, as do those of AMAE . This metric has
been chosen given the imbalanced nature of the problem considered.
4.3 Results
The results obtained for all the methods can be seen in Table 1 for Acc and Table
2 for AMAE, from which several conclusions can be drawn. Note that the results
without feature selection are also included in the Tables for the three datasets.
Firstly, the performance of the base algorithm (i.e. with no feature selection) can
be improved in some cases (e.g. in SW-5c), and there are reduction levels where
the reduced datasets perform relatively similar to the base ones (e.g. with 30% of
features), which is interesting for model interpretability purposes and to reduce
Table 1. Acc mean and standard deviation (Mean ±SD) obtained by all the method-
ologies compared as a function of the percentage of features selected.
Perc. of features SW-3c, result without FS: 68.99 ±0.53
10% 46.31 ±2.55 49.02 ±3.75 49 .41 ±2.29 45.11 ±4.67 55.52 ±4.55
20% 56.83 ±4.45 56 .76 ±6.28 53.14 ±1.41 49.59 ±2.46 49.99 ±2.58
30% 60.25 ±0.21 60 .35 ±0.09 52.36 ±1.32 60.14 ±0.57 65.26 ±0.67
40% 60 .34 ±0.27 60.33 ±0.18 56.69 ±1.31 62.14 ±0.83 65.01 ±0.80
50% 60.40 ±0.07 60 .41 ±0.05 60.32 ±0.12 64.21 ±0.71 64.22 ±0.65
60% 59.93 ±1.00 60.05 ±0.82 59.53 ±1.01 66 .48 ±0.66 66.49 ±0.63
70% 60.36 ±0.51 60.77 ±0.74 61.74 ±0.69 67.37 ±0.62 67 .36 ±0.68
80% 62.91 ±0.52 63.46 ±0.68 64.27 ±0.69 68.39 ±0.51 68 .28 ±0.48
90% 66.09 ±0.52 66.18 ±0.51 66.98 ±0.64 68 .87 ±0.50 68.89 ±0.54
Perc. of features SW-5c, result without FS: 50.22 ±0.49
10% 30 .38 ±3.49 32.41 ±4.24 26.62 ±9.84 22.10 ±3.26 22.73 ±3.16
20% 38.15 ±10.35 44.75 ±7.18 29.63 ±3.69 42 .32 ±3.73 42.28 ±3.73
30% 48 .67 ±0.25 48.40 ±2.24 38.17 ±1.67 48.84 ±0.06 48.84 ±0.06
40% 48.82 ±0.07 48 .84 ±0.04 42.84 ±0.83 48.86 ±0.02 48.86 ±0.02
50% 48 .84 ±0.01 48.85 ±0.01 47.65 ±0.50 48.69 ±0.42 48.67 ±0.43
60% 48.85 ±0.00 48.85 ±0.00 48.85 ±0.01 49 .72 ±0.59 50.74 ±0.66
70% 48.82 ±0.18 48.84 ±0.05 48.78 ±0.18 50 .80 ±0.41 51.10 ±0.40
80% 48.88 ±0.22 49.05 ±0.35 49.51 ±0.34 50 .88 ±0.31 51.48 ±0.22
90% 49.88 ±0.33 49.94 ±0.30 50.21 ±0.27 50 .42 ±0.23 51.08 ±0.20
Perc. of features SW-11c, result without FS: 30.09 ±0.34
10% 10.85 ±3.72 15.82 ±3.27 6.70 ±4.80 14.35 ±2.34 14 .60 ±2.56
20% 17.97 ±6.07 24 .58 ±6.61 13.06 ±2.86 22.52 ±3.10 27.88 ±4.08
30% 30 .49 ±0.30 29.80 ±3.12 19.30 ±1.07 30.71 ±0.08 30.33 ±1.14
40% 30 .59 ±0.41 30.60 ±0.49 21.83 ±1.46 29.07 ±2.37 27.78 ±0.47
50% 30 .51 ±0.28 30.61 ±0.20 30.01 ±1.35 27.54 ±0.86 29.38 ±0.41
60% 30.73 ±0.03 30.73 ±0.04 30 .35 ±1.32 28.77 ±0.57 30.18 ±0.36
70% 29 .47 ±1.99 28.51 ±2.01 27.39 ±1.26 29.41 ±0.47 30.37 ±0.31
80% 27.66 ±0.37 28.13 ±0.48 28.53 ±0.51 29 .79 ±0.34 30.20 ±0.35
90% 29.01 ±0.28 29.21 ±0.27 29.44 ±0.40 29 .98 ±0.35 30.11 ±0.31
Average 43.777 44.639 41.233 44.977 45.838
Ranking 3.519 2.778 3.981 2.759 1.963
Friedman’s test Confidence interval C0= (0,F(α=0.05) = 2.459). F-val.Acc : 8.278 /C0.
The best performing method is in b old face and the second one in italics.
computational time. Secondly, the results are in general promising (considering
the difficulty of the problem). Finally, the proposed technique performs better
than the nominal counterpart (specially when the Hausdorff distance is used).
To quantify whether a statistical difference exists among the algorithms
compared, a procedure is employed to compare multiple classifiers in multiple
datasets [17]. Tables 1 and 2 also shows the result of applying the non-parametric
statistical Friedman’s test (for a significance level of α= 0.05) to the mean Acc
and AMAE rankings. The test rejected the null-hypothesis that all of the algo-
rithms perform similarly in mean ranking for both metrics.
On the basis of this rejection and following the guidelines in [17], we consider
the best performing method as control method for the following test. We compare
this method to the rest according to their rankings. It has been noted that the
approach of comparing all classifiers to each other in a post-hoc test is not as
sensitive as the approach of comparing all classifiers to a given classifier (a control
method). One approach to this latter type of comparison is the Holm’s test. The
test statistics for comparing the i-th and j-th method using this procedure is:
Table 2. AM AE mean and standard deviation (Mean ±SD) obtained by all the
methodologies compared as a function of the percentage of features selected.
Perc. of features SW-3c, result without FS: 0.581 ±0.010
10% 0.877 ±0.037 0.891 ±0.039 0.870 ±0.036 0.825 ±0.049 0.780 ±0.054
20% 0.947 ±0.057 0.955 ±0.084 0.756 ±0.020 0.757 ±0.035 0.758 ±0.052
30% 0.996 ±0.006 0.999 ±0.002 0.731 ±0.019 0.917 ±0.105 0.606 ±0.108
40% 0.997 ±0.009 0.997 ±0.007 0.913 ±0.046 0.647 ±0.014 0.606 ±0.014
50% 0.997 ±0.003 0.999 ±0.002 0.997 ±0.006 0.631 ±0.009 0.630 ±0.009
60% 0.952 ±0.098 0.950 ±0.102 0.806 ±0.140 0.610 ±0.007 0.610 ±0.008
70% 0.777 ±0.114 0.730 ±0.075 0.677 ±0.020 0.600 ±0.007 0.600 ±0.008
80% 0.678 ±0.016 0.668 ±0.016 0.645 ±0.017 0.588 ±0.009 0.589 ±0.009
90% 0.630 ±0.014 0.629 ±0.015 0.609 ±0.014 0.581 ±0.008 0.581 ±0.008
Perc. of features SW-5c, result without FS: 1.294 ±0.052
10% 1.619 ±0.098 1.590 ±0.107 1.668 ±0.192 1.678 ±0.116 1.668 ±0.097
20% 1.490 ±0.199 1.460 ±0.168 1.486 ±0.085 1.349 ±0.080 1.347 ±0.079
30% 1.401 ±0.001 1.413 ±0.070 1.404 ±0.023 1.399 ±0.001 1.399 ±0.001
40% 1.400 ±0.001 1.400 ±0.000 1.410 ±0.014 1.400 ±0.000 1.400 ±0.000
50% 1.400 ±0.000 1.400 ±0.000 1.401 ±0.005 1.356 ±0.100 1.339 ±0.112
60% 1.400 ±0.000 1.400 ±0.000 1.400 ±0.000 1.160 ±0.049 1.137 ±0.066
70% 1.396 ±0.022 1.396 ±0.023 1.370 ±0.062 1.192 ±0.021 1.145 ±0.023
80% 1.358 ±0.050 1.336 ±0.054 1.280 ±0.045 1.249 ±0.017 1.181 ±0.017
90% 1.304 ±0.028 1.297 ±0.022 1.284 ±0.016 1.294 ±0.010 1.232 ±0.010
Perc. of features SW-11c, result without FS: 2.884 ±0.015
10% 3.986 ±0.353 3.774 ±0.255 4.530 ±0.634 3.381 ±0.229 3.384 ±0.200
20% 3.670 ±0.425 3.549 ±0.146 3.696 ±0.333 3.143 ±0.145 3.404 ±0.215
30% 3.535 ±0.011 3.514 ±0.111 3.298 ±0.150 3.546 ±0.001 3.494 ±0.143
40% 3.535 ±0.025 3.537 ±0.030 3.246 ±0.087 3.264 ±0.403 2.656 ±0.049
50% 3.528 ±0.018 3.534 ±0.017 3.520 ±0.076 2.767 ±0.157 2.685 ±0.040
60% 3.543 ±0.003 3.542 ±0.004 3.497 ±0.167 2.765 ±0.047 2.739 ±0.030
70% 3.386 ±0.248 3.222 ±0.289 2.943 ±0.212 2.802 ±0.035 2.799 ±0.025
80% 2.976 ±0.024 2.929 ±0.036 2.861 ±0.046 2.843 ±0.021 2.848 ±0.020
90% 2.913 ±0.022 2.905 ±0.023 2.875 ±0.028 2.870 ±0.018 2.867 ±0.016
Average 1.914 1.889 1.858 1.712 1.647
Ranking 4.185 3.852 3.093 2.370 1.500
Friedman’s test Confidence interval C0= (0, F(α=0.05) = 2.459). F-val.AMAE : 23.859 /C0.
The best performing method is in b old face and the second one in italics.
,where Jis the number of algorithms, Tis the number of datasets
and Riis the mean ranking of the i-th method. The zvalue is used to find the
corresponding probability from the table of the normal distribution, which is
then compared with an appropriate level of significance α. Holm’s test adjusts
the value for αin order to compensate for multiple comparisons. This is done
in a step-up procedure that sequentially tests the hypotheses ordered by their
significance. We will denote the ordered p-values by p1, p2, . . . , pqso that p1
p2. . . pq. Holm’s test compares each piwith α
Holm =α/(Ji), starting
from the most significant pvalue. If p1is below α/(J1), the corresponding
hypothesis is rejected, and we allow to compare p2with α/(J2). If the second
hypothesis is rejected, the test proceeds with the third, and so on.
Table 3 shows the result of the Holm’s test when comparing the best per-
forming technique (i.e. OFSHausdorff) to the rest of algorithms. It can be seen
that this method outperforms the rest of techniques for AM AE and specifically
OFSMean and NFS for Acc, when α= 0.05. This result validates the Hausdorff
distance in this context and shows that the ordinal nature of the data should
be considered when performing feature selection. The results not only improve
when considering ordinal measures (i.e. AMAE) but also nominal ones (such as
Acc), meaning that the method can benefit from the order constraint introduced.
Table 3. Results of the Holm procedure using OFSHausdorff as control method: cor-
rected αvalues, compared method and p-values, ordered by comparisons (i).
Control alg.: OFSHausdorff Acc
i α
0.05 Method pi
1 0.01250 OFSMean 0.00000++
2 0.01667 NFS 0.00003++
3 0.02500 OFS 0.05830
4 0.05000 OFSMin 0.06425
Control alg.: OFSHausdorff AM AE
i α
0.05 Method pi
1 0.01250 NFS 0.00000++
2 0.01667 OFS 0.00000++
3 0.02500 OFSMean 0.00002++
4 0.05000 OFSMin 0.04312++
Win (++) with statistical significant difference for α= 0.05
4.4 Discussion
As stated before, the authors consider that a percentage of 30% features repre-
sents a good option, since it allows to have a relatively accurate model with an
increased interpretability (70% of the variables are removed). In this section, we
examine the variables (associated to this 30%) that are of largest importance to
the characterisation of happiness, considering the version of the dataset with 3
classes and the OFSHausdorff method. Note that the feature selection method
does not consider aggregated sets of features, and therefore is not able to discover
redundant variables of interactions between them. Since we considered 30 dif-
ferent train data partitions, we have 30 different results. The following variables
have been selected at least in 15 models: variables related to the country (mean-
ing that there could be other factors of vital importance to the characterisation
of well-being, such as the state of economy of the country, the weather, etc.),
gender, variables related to love relationships (single, legally married, living with
partner, etc.), general level of health, whether the person is daily hampered by
illness or if he/she belongs to a minority ethnic, the number of times the person
felt depressed previous week, the quality of the sleep, whether the person gives
importance to be rich and to have expensive belongings, religiosity and whether
the person would allow entry of immigrants from poorer countries.
Several conclusions can be drawn from these results. Firstly, the environment
in which a person lives significantly influences their well-being and so do other
demographic variables (e.g. familiar composition, gender or belonging to a mi-
nority ethnic). Secondly, both physical and mental health have an impact on
subjective well-being. Finally, the group of variables related to personality and
opinions also play a vital role in the classification of happiness.
The variables selected when considering 70% of features were also examined
in order to study variables that were selected by very few models. Note from
Table 1 that 70% represents also an interesting threshold, as it is the point
from which the performance usually starts to degrade. In this case, the range of
variables present in at least in 15 models includes additional variables concern-
ing: social well-being (feeling of loneliness and inability to get going, number of
people to discuss personal matters with, involvement in social activities and con-
flicts at home when growing up), health and habits (diet/sport/alcohol/smoking
behaviour, inability to get medical consultation or treatment), daily activity
(specifically the main activity, the number of hours watching TV per week and
whether the person improved their knowledge/skills in the last year), other de-
mographic factors (such as the domicile, the type of contract or the familiar com-
position), community well-being (interest in politics, closeness to country, satis-
faction with health and education in country) and, finally, personality (whether
it is important to be modest and to seek for adventures, whether the person
thinks that other people try to take advantage of them and the feeling of safety
walking alone in local area at night). Conversely, there is a set of variables that
have been selected in very few models (in this case we consider variables selected
in less than 10 models). These variables are the following: age, weight, number of
hours working, people responsible for at job, financial difficulties when growing
up, years of education, number of people living with, satisfaction with govern-
ment and economy (as opposed to health and education that were selected as
relevant), whether politicians care what people think and the number of hours
helping others (family, friends or neighbours).
5 Conclusions and future work
This paper presents a feature selection strategy for classification problems where
the dependent variable follows a natural order. We construct a dataset for pre-
dicting subjective well-being across different European countries that includes
56 variables of different components of happiness. The results show that there
are some factors, such as the environment where the person lives, the physical
and mental health and the personality, that are of great influence to subjective
well-being. Moreover, the performance of the proposed method is competitive
against its nominal counterpart, which demonstrates the necessity of developing
more specific techniques for domains such as the ordinal classification one.
Future research comprises including other countries (e.g. developing ones),
performing a sensitivity analysis, comparing the features selected for the different
versions of the dataset and extending the proposed methodology to consider
aggregated sets of features.
1. Linley, P.A., Maltby, J., Wood, A.M., Osborne, G., Hurling, R.: Measuring happi-
ness: The higher order factor structure of subjective and psychological well-being
measures. Personality and Individual Differences 47(8) (2009) 878 – 884
2. Diener, E.: Subjective well-being: The science of happiness and a proposal for a
national index. American Psychologist 55(1) (2000) 34–43
3. Self, A., Thomas, J., Randall, C.: Measuring national well-being: Life in the uk
(2012) Last access: 8 dec 2015.
4. Keyes, C.L., Shmotkin, D., Ryff, C.D.: Optimizing well-being: the empirical en-
counter of two traditions. Journal of personality and social psychology 82(6) (2002)
5. Guti´errez, P.A., P´erez-Ortiz, M., S´anchez-Monedero, J., Fern´andez-Navarro, F.,
Herv´as-Mart´ınez, C.: Ordinal regression methods: Survey and experimental study.
Knowledge and Data Engineering, IEEE Transactions on 28(1) (Jan 2016) 127–146
6. Gu, Q., Li, Z., Han, J.: Generalized fisher score for feature selection. CoRR
abs/1202.3725 (2012)
7. Bixter, M.T.: Happiness, political orientation, and religiosity. Personality and
Individual Differences 72 (2015) 7 – 11
8. Hills, P., Argyle, M.: The Oxford Happiness Questionnaire: a compact scale for the
measurement of psychological well-being. Personality and Individual Differences
33(7) (November 2002) 1073–1082
9. Chiu, D., Wong, A.: Synthesizing knowledge: A cluster analysis approach using
event-covering. IEEE Transactions on Systems and Man and Cybernetics and Part
B16(2) (1986) 251–259
10. Luengo, J., Garc´ıa, S., Herrera, F.: On the choice of the best imputation methods
for missing values considering three groups of classification methods. Knowledge
and Information Systems 32(1) (2012) 77–108
11. erez-Ortiz, M., Guti´errez, P.A., Herv´as-Mart´ınez, C.: Projection-based ensemble
learning for ordinal regression. Cybernetics, IEEE Transactions on 44(5) (2014)
12. Baccianella, S., Esuli, A., Sebastiani, F.: Feature selection for ordinal text classi-
fication. Neural Comput. 26(3) (March 2014) 557–591
13. Mukras, R., Wiratunga, N., Lothian, R., Chakraborti, S., Harper, D.: Infor-
mation gain feature selection for ordinal text classification using probability re-
distribution. In: the IJCAI’07 Workshop on Text Mining and Link Analysis, Hy-
derabad, IN (2007)
14. Eiter, T., Mannila, H.: Distance measures for point sets and their computation.
Acta Informatica 34 (1997) 103–133
15. Sun, B.Y., Li, J., Wu, D.D., Zhang, X.M., Li, W.B.: Kernel discriminant learning
for ordinal regression. IEEE Transactions on Knowledge and Data Engineering 22
(2010) 906–910
16. Baccianella, S., Esuli, A., Sebastiani, F.: Evaluation measures for ordinal regres-
sion. In: Proceedings of the Ninth International Conference on Intelligent Systems
Design and Applications (ISDA 09), Pisa, Italy (December, 2009 2009)
17. Demsar, J.: Statistical comparisons of classifiers over multiple data sets. Journal
of Machine Learning Research 7(2006) 1–30
... The following is the common opinion about CURE-SMOTE algorithm: use CURE to group the minor class's samples from the actual samples, take away the noise and outliers, and at last, randomly generate artificial samples between the representative and center point. The following is the CURE-SMOTE algorithm's execution steps [18,19]. ...
Full-text available
Access Control Models (ACM) must secure the communication system of devices in Internet of Vehicles (IoV) under cloud computing architecture. Existing ACM, on the other hand, struggles to determine the right granularity of permissions when dealing with vast numbers of data in the IoV. Furthermore, IoV is vulnerable to attacks, as attackers can readily exploit existing flaws. Due to insufficient or inefficient ACM, some attacks may succeed. As a result, the authentication mechanism must be reinforced as much as possible using cutting-edge ACM. Methods have been applied to Decisions Making System (DMS) about who has access to what in open distributed information systems like big data, the Internet of Things (IoT), and the cloud experience performance issues because of the number and complexity of the rules and regulations governing who has significant exposure to what. The reasonably significant Access Control (AC) time operational costs have a negative impact on the regular functioning of business services as a consequence. This paper presents a framework for an efficient SDN-involved Dynamic Access system based on AdaBoost (SDNDAAB) model. The challenges related to the ACM is changed by this proposed model into a Binary Classification Problem (BCP) that either permit access permission or deny them. So, apart from providing dynamic support for the AC’s efficient execution amidst IoV, the AdaBoost algorithm also supports the disseminated application of the decision engine via a Software Defined Network (SDN) controller for predicting the AC. The results show that the proposed model supports better permission decision accuracy than the other models.
... Fisher Score (FS) [12]: It is calculated as the ratio of between scatter ( ...
... In order to eliminate the curse of dimensionality problem in high-dimensional dataset, top-100 features were selected by applying the Fisher scoring feature selection strategy to original feature set in experimental analysis. In the study conducted by Pe rez-Ortiz et al. [5] subjective well-being factors were successfully extracted by Fisher scoring method, using data containing 56 different components of the happiness in European countries. In another study conducted by Gu et al. the generalized Fisher scoring method, which maximizes the lower bound of the traditional Fisher score, was presented and experiments showed that the proposed method outperformed many state-of-the-art feature selection methods [6]. ...
Deep belief network (DBN) is deep neural network structure consisting of a collection of restricted Boltzmann machine (RBM). RBM is two-layered simple neural networks which are formed by a visible and hidden layer, respectively. Each visible layer receives a lower-level feature set learned by previous RBM and passes it through to top layers turning them into a more complex feature structure. In this study, the proposed method is to feed the training parameters learned by DBN to multilayer perceptron as initial weights instead of starting them from random points. The obtained results on the bioinformatics cancer dataset show that using initial weights trained by DBN causes more successful classification results than starting from random parameters. The test accuracy using proposed method increased from 77.27 to 95.45%.
... -Ordinal Fisher (OF) score [20] is an ordinal adaptation of the Fisher score [7]. This measure gives higher penalisation for distant classes in the ordinal scale, therefore, distant classes should be associated with higher distances. ...
Time series ordinal classification is one of the less studied problems in time series data mining. This problem consists in classifying time series with labels that show a natural order between them. In this paper, an approach is proposed based on the Shapelet Transform (ST) specifically adapted to ordinal classification. ST consists of two different steps: 1) the shapelet extraction procedure and its evaluation; and 2) the classifier learning using the transformed dataset. In this way, regarding the first step, 3 ordinal shapelet quality measures are proposed to assess the shapelets extracted, and, for the second step, an ordinal classifier is applied once the transformed dataset has been constructed. An empirical evaluation is carried out, considering 7 ordinal datasets from the UEA & UCR Time Series Classification (TSC) repository. The results show that a support vector ordinal classifier applied to the ST using the Pearson’s correlation coefficient (\(R^2\)) is the combination achieving the best results in terms of two evaluation metrics: accuracy and average mean absolute error. A final comparison against three of the most popular and competitive nominal TSC techniques is performed, demonstrating that ordinal approaches can achieve higher performances even in terms of accuracy.
... The first of these measures is based on the Fisher score [19], commonly used for feature selection. A reformulation of this score has been made in order to adapt it to ordinal classification, known as Ordinal Fisher (OF) [20] score. This reformulation is based on the inclusion of higher costs for distant classes, i.e. the cost depends on the distance between the shapelet class and the class of the time series being compared. ...
... The feature selection techniques are mostly categorized into two categories. The relief [7], Fisher score [8], Chi-squared score [9], correlation-based feature selection [10], and fast correlation-based filter [11] are considered as supervised techniques. The most common unsupervised methods are the mean [12], variance [12], skewness [13], kurtosis [13], mean absolute difference [13], multi-cluster feature selection [14], laplacian score [15], dispersion ratio [13], laplacian score combined with distance-based entropy [15], mutivariate and univariate statistical methods such as squared prediction error statistic [16], T 2 statistic [17], ϕ statistic [18] and generalized likelihood ratio test [19]. ...
The random forest (RF) classifier, which is a combination of tree predictors, is one of the most powerful classification algorithms that has been recently applied for fault detection and diagnosis (FDD) of industrial processes. However, RF is still suffering from some limitations such as the noncorrelation between variables. These limitations are due to the direct use of variables measured at nodes and therefore the only use of static information from the process data. Thus, this article proposes two enhanced RF classifiers, namely the Euclidean distance based reduced kernel RF (RKRF-ED) and K-means clustering based reduced kernel RF (RKRF-Kmeans), for FDD. Based on the kernel principal component analysis, the proposed classifiers consist of two main stages: feature extraction and selection, and fault classification. In the first stage, the number of observations in the training data set is reduced using two methods: the first method consists of using the Euclidean distance as dissimilarity metric so that only one measurement is kept in case of redundancy between samples. The second method aims at reducing the amount of the training data based on the K-means clustering technique. Once the characteristics of the process are extracted, the most sensitive features are selected. During the second phase, the selected features are fed to an RF classifier. An emulated grid-connected PV system is used to validate the performance of the proposed RKRF-ED and RKRF-Kmeans classifiers. The presented results confirm the high classification accuracy of the developed techniques with low computation time.
... Using the well-known Fisher ratio concept, the Fisher score is used to select required features and a heuristic policy is deployed to determine a score for features [23]. The advantages are  To identify the relevant features for any specific problem  Reduces the size of the problem and computer storage  Reduce the computation time also to improve the quality of prediction  To improve the classifier by removing the irrelevant features and noise Let the average and standard deviation of the k -th class and i-th function (and μ i , σ i the mean and standard deviation of the entire dataset for the i-th function) be μ i k and σ i k . ...
Full-text available
This paper is aimed to analyze the feature selection process based on different statistical methods viz., Correlation, Gain Ratio, Information gain, OneR, Chi-square MapReduce model, Fisher's exact test for agricultural data. During the recent past, Fishers exact test was commonly used for feature selection process. However, it supports only for small data set. To handle large data set, the Chi square, one of the most popular statistical methods is used. But, it also finds irrelevant data and thus resultant accuracy is not as expected. As a novelty, Fisher's exact test is combined with Map Reduce model to handle large data set. In addition, the simulation outcome proves that proposed fisher's exact test finds the significant attributes with more accurate and reduced time complexity when compared to other existing methods.
... The feature selection techniques are mostly categorized into two categories. The relief [7], Fisher score [8], Chi-squared score [9], correlation-based feature selection [10], and fast correlation-based filter [11] are considered as supervised techniques. The most common unsupervised methods are the mean [12], variance [12], skewness [13], kurtosis [13], mean absolute difference [13], multi-cluster feature selection [14], laplacian score [15], dispersion ratio [13], laplacian score combined with distance-based entropy [15], mutivariate and univariate statistical methods such as squared prediction error statistic [16], T 2 statistic [17], ϕ statistic [18] and generalized likelihood ratio test [19]. ...
Conference Paper
Process monitoring is an essential part of industrial systems. It requires higher product quality and safety operations. Therefore, a nonlinear data-driven approach based reduced KPCA (RKPCA) for statistical monitoring of industrial processes is developed. RKPCA is a novel machine learning tool which merges dimensionality reduction and supervised learning. The use of classical KPCA for modeling and monitoring purposes can impose a high computational load when a large number of measurements are recorded. The main idea of the proposed RKPCA approach is to reduce the number of observations (samples) in the data matrix using the Euclidean distance between samples as dissimilarity metric so that only one observation is kept in case of redundancy. The Tennessee Eastman Process (TEP) is used to evaluate the fault detection abilities of the proposed RKPCA technique. The performance of the proposed method is evaluated and compared to the classical KPCA in terms of false alarms rates (FAR), missed detection rates (MDR) and computation times (CT).
Dental healthcare providers need to examine a large number of panoramic X-ray images every day. It is quite time consuming, tedious, and error-prone job. The examination quality is also directly related to the experience and the personal factors, i.e., stress, fatigue, etc., of the dental care providers. To assist them handling this problem, a residual network-based deep learning technique, i.e., faster R-CNN technique, is proposed in this study. Two kinds of residual networks, i.e., ResNet-50 and ResNet-101, are used as the base network of faster R-CNN separately. A modified version of Palmer notation (PN) system is proposed in this research for numbering the teeth. The modified Palmer notation (MPN) system does not use any notation like PN system. The MPN system is proposed for mainly three reasons: (i) teeth are divided into total eight categories, and to keep this similarity, a new numbering system is proposed that has the same number of category, (ii) 8-category MPN system is less complex to implement than 32-category universal tooth numbering (UTN) system, and with some preprocessing steps, MPN system can be converted into 32-category UTN system, and finally (iii) for the convenience of the dentist, i.e., it is more feasible to utilize 8-category MPN system than 32-category UTN system. Total 900 dental X-ray images were used as training data, while 100 images were used as test data. The method achieved 0.963 and 0.965 mean average precision (mAP) for ResNet-50 and ResNet-101, respectively. The obtained results demonstrate the effectiveness of the proposed method and satisfy the condition of clinical implementation. Therefore, the method can be considered as a useful and reliable tool to assist the dental care providers in dentistry.
Full-text available
Ordinal regression problems are those machine learning problems where the objective is to classify patterns using a categorical scale which shows a natural order between the labels. Many real-world applications present this labelling structure and that has increased the number of methods and algorithms developed over the last years in this field. Although ordinal regression can be faced using standard nominal classification techniques, there are several algorithms which can specifically benefit from the ordering information. Therefore, this paper is aimed at reviewing the state of the art on these techniques and proposing a taxonomy based on how the models are constructed to take the order into account. Furthermore, a thorough experimental study is proposed to check if the use of the order information improves the performance of the models obtained, considering some of the approaches within the taxonomy. The results confirm that ordering information benefits ordinal models improving their accuracy and the closeness of the predictions to actual targets in the ordinal scale.
Full-text available
Ordinal classification (also known as ordinal regression) is a supervised learning task that consists of estimating the rating of a data item on a fixed, discrete rating scale. This problem is receiving increased attention from the sentiment analysis and opinion mining community due to the importance of automatically rating large amounts of product review data in digital form. As in other supervised learning tasks such as binary or multiclass classification, feature selection is often needed in order to improve efficiency and avoid overfitting. However, although feature selection has been extensively studied for other classification tasks, it has not for ordinal classification. In this letter, we present six novel feature selection methods that we have specifically devised for ordinal classification and test them on two data sets of product review data against three methods previously known from the literature, using two learning algorithms from the support vector regression tradition. The experimental results show that all six proposed metrics largely outperform all three baseline techniques (and are more stable than these others by an order of magnitude), on both data sets and for both learning algorithms.
Full-text available
This paper looks at feature selection for ordinal text clas-sification. Typical applications are sentiment and opinion classification, where classes have relationships based on an ordinal scale. We show that standard feature selection using Information Gain (IG) fails to identify discriminatory features, particularly when they are distributed over mul-tiple ordinal classes. This is because inter-class similarity, implicit in the ordinal scale, is not exploited during feature selection. The Probability Re-distribution Procedure (PRP), introduced in this paper, explicates inter-class similarity by revising feature distributions. It aims to influ-ence feature selection by improving the ranking of features that are dis-tributed over similar classes, relative to those distributed over dissimilar classes. Evaluations on three datasets illustrate that the PRP helps select features that result in significant improvements on classifier performance. Future work will focus on automated acquisition of inter-class similarity knowledge, with the aim of generalising the PRP for a wider class of problems.
Full-text available
The classification of patterns into naturally ordered labels is referred to as ordinal regression. This paper proposes an ensemble methodology specifically adapted to this type of problem, which is based on computing different classification tasks through the formulation of different order hypotheses. Every single model is trained in order to distinguish between one given class (k) and all the remaining ones, while grouping them in those classes with a rank lower than k, and those with a rank higher than k. Therefore, it can be considered as a reformulation of the well-known one-versus-all scheme. The base algorithm for the ensemble could be any threshold (or even probabilistic) method, such as the ones selected in this paper: kernel discriminant analysis, support vector machines and logistic regression (LR) (all reformulated to deal with ordinal regression problems). The method is seen to be competitive when compared with other state-of-the-art methodologies (both ordinal and nominal), by using six measures and a total of 15 ordinal datasets. Furthermore, an additional set of experiments is used to study the potential scalability and interpretability of the proposed method when using LR as base methodology for the ensemble.
Full-text available
In real-life data, information is frequently lost in data mining, caused by the presence of missing values in attributes. Several schemes have been studied to overcome the drawbacks produced by missing values in data mining tasks; one of the most well known is based on preprocessing, formerly known as imputation. In this work, we focus on a classification task with twenty-three classification methods and fourteen different imputation approaches to missing values treatment that are presented and analyzed. The analysis involves a group-based approach, in which we distinguish between three different categories of classification methods. Each category behaves differently, and the evidence obtained shows that the use of determined missing values imputation methods could improve the accuracy obtained for these methods. In this study, the convenience of using imputation methods for preprocessing data sets with missing values is stated. The analysis suggests that the use of particular imputation methods conditioned to the groups is required. KeywordsApproximate models–Classification–Imputation–Rule induction learning–Lazy learning–Missing values–Single imputation
Full-text available
The nature and structure of well-being is a topic that has garnered increasing interest with the emergence of positive psychology. Limited research to date suggests two separate but related factors of subjective well-being and psychological well-being. Subjective well-being comprises an affective component of the balance between positive and negative affect, together with a cognitive component of judgments about one’s life satisfaction. Psychological well-being is conceptualised as having six components, including positive relations with others, autonomy, environmental mastery, self-acceptance, purpose in life and personal growth. In the current study, we used exploratory factor analysis and confirmatory factor analysis to examine the higher order factor structure of subjective and psychological well-being in a series of large UK samples. Analyses showed that subjective well-being and psychological well-being loaded separately onto two independent but related factors, consistent with previous research. Further, we demonstrated that these loadings did not vary according to gender, age or ethnicity, providing further support for the robustness of this higher order factor structure. The discussion locates these findings in context and explores future research directions on the associations between subjective and psychological well-being over time.
Full-text available
Fisher score is one of the most widely used supervised feature selection methods. However, it selects each feature independently according to their scores under the Fisher criterion, which leads to a suboptimal subset of features. In this paper, we present a generalized Fisher score to jointly select features. It aims at finding an subset of features, which maximize the lower bound of traditional Fisher score. The resulting feature selection problem is a mixed integer programming, which can be reformulated as a quadratically constrained linear programming (QCLP). It is solved by cutting plane algorithm, in each iteration of which a multiple kernel learning problem is solved alternatively by multivariate ridge regression and projected gradient descent. Experiments on benchmark data sets indicate that the proposed method outperforms Fisher score as well as many other state-of-the-art feature selection methods.
Previous research has focused on how happiness is independently associated with political orientation and religiosity. The current study instead explored how political orientation and religiosity interact in establishing levels of happiness. Data from both the 2012 General Social Survey and the 2005 World Values Survey were used. Results from both data sets support prior research by showing a positive association between happiness and both political conservatism and religiosity. Importantly, it was found that political conservatism and religiosity interact in predicting happiness levels. Specifically, the current results suggest that religiosity has a greater effect on happiness for more politically conservative individuals compared to more politically liberal individuals.
An improved instrument, the Oxford Happiness Questionnaire (OHQ), has been derived from the Oxford Happiness Inventory, (OHI). The OHI comprises 29 items, each involving the selection of one of four options that are different for each item. The OHQ includes similar items to those of the OHI, each presented as a single statement which can be endorsed on a uniform six-point Likert scale. The revised instrument is compact, easy to administer and allows endorsements over an extended range. When tested against the OHI, the validity of the OHQ was satisfactory and the associations between the scales and a battery of personality variables known to be associated with well-being, were stronger for the OHQ than for the OHI. Although parallel factor analyses of OHI and the OHQ produced virtually identical statistical results, the solution for the OHQ could not be interpreted. The previously reported factorisability of the OHI may owe more to the way the items are formatted and presented, than to the nature of the items themselves. Sequential orthogonal factor analyses of the OHQ identified a single higher order factor, which suggests that the construct of well-being it measures is uni-dimensional. Discriminant analysis has been employed to produce a short-form version of the OHQ with eight items.