See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/313867740

A review of missing values handling methods on time-series data

Conference Paper · October 2016

DOI: 10.1109/ICITSI.2016.7858189

CITATIONS READS

19 9,160

4 authors, including:

Irfan Pratama Adhistya Erna Permanasari

Universitas Gadjah Mada Universitas Gadjah Mada
5 PUBLICATIONS   43 CITATIONS    84 PUBLICATIONS   312 CITATIONS   

SEE PROFILE SEE PROFILE

Rini Indrayani
Universitas AMIKOM Yogyakarta
6 PUBLICATIONS   37 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Study Of Gaze-based Object Selection For Spontaneous Interaction View project

FUKURO the Wheeled-Robosoccer of Universitas Gadjah Mada View project

All content following this page was uploaded by Irfan Pratama on 05 April 2018.

The user has requested enhancement of the downloaded file.

 2016 International Conference on Information Technology Systems and Innovation (ICITSI)
Bandung – Bali, October 24 – 27, 2016
ISBN: 978-1-5090-2449-0

A Review of Missing Values Handling Methods

on Time-Series Data
Irfan Pratama1, Adhistya Erna Permanasari2, Igi Ardiyanto3, Rini Indrayani4
Department of Electrical Engineering and Information Technology
Universitas Gadjah Mada
Yogyakarta, Indonesia
1
[email protected], [email protected], [email protected], [email protected]

Abstract — Missing values becomes one of the problems because of an observation devices breakdown. While in an
that frequently occur in the data observation or data opinion survey or an interview, missing values caused by
recording process. The needs of data completeness of the respondents who decline to answer or complete the survey,
observation data for the uses of advanced analysis becomes or even because insufficient survey observation’s stuff
important to be solved. Conventional method such as [2,3]. Both examples stated above are trully unexpected
mean and mode imputation, deletion, and other methods occurrence. In other word, the data are expected to be observed
are not good enough to handle missing values as those (censor failure cases), yet it is missing due to the setbacks.
method can caused bias to the data. Estimation or Missing values is probably can not be evaded as it is happened
imputation to the missing data with the values produced unexpectedly. Time series data are become one of some data
by some procedures or algorithms can be the best possible that most likely to have missing values in it.
solution to minimized the bias effect of the conventional Time series is a data that observed in regular interval of
method of the data. So that at last, the data will be times [4]. Time series form can found as a continous data or
completed and ready to use for another step of analysis or discrete because observed in a regular interval of time
data mining. In this paper, we will explain and describe (example: daily, monthly, annualy). Some examples of
several previous studies about missing values handling continous and discrete time series data are sinusoidal signal
methods or approach on time series data. This paper also which is continous time series data, while daily stock prices,
discuss some plausible option of methods to estimate temperature data are a discrete data.
missing values to be used by other researchers in this field Data measurements are conducted several times with
of study. The discussion’s aim is to help them to figure out different condition, and sometimes, missing data occured due
what method is commonly used now along with its to several problem that are known as the “missingness
advantages and drawbacks. mechanism” [5,6,7]:
1) Missing completely at random (MCAR): a variable
Keywords — missing values, estimation technique, mean is missing completely at random if the probability of
imputation, deletion, time series. missingness is the same for all units, or in other words the
is no dependencies of the missingness probability related to
I. INTRODUCTION
the variable itself [7,8,9]
A pack of data or dataset can be used to obtain a certain 2) Missing at random (MAR): a variable is missing
specific information which can give a new knowledge. It can at random if the probability of missingness is depending
be obtained as classification, trend, pattern, etc [1]. Observed only on available information.
data can be obtained by several methods, such as censor record 3) Not missing at random (NMAR): the missingness
(automaticly and continously) or by frequent observation probability is depending on the variable itself.
(survey, medical record, etc). Each of analysis process consists Diffferent way of data measurement will emerging
of several steps, one of them is preprocessing. Preprocessing is different mechanism assumption, for example: censor
a step or phase to identify, selection, or problem handling of recording failure will be treated as a missing completely at
the data. Missing values handling is included in the random since it has no dependencies to the data that are
preprocessing step. Missing values is the most common missing, but in a survey, a respondent who refused to report
problem that occured in data that is obtained by observation his income report will be treated as not missing at random
and censor recording. since the surveyors already expected that such certain data
Missing data occured because of several problems, such are unlikely to be obtained easily. Several methods has been
as technical fault or human errors (the object of observation introduced to solve missing values according to its missing
did not give sufficient data to the observer). Some real world mechanism or even a general solution for any missing values
cases such as industrial experiment, missing values occured mechanism from the conventional one to the more modern

978-1-5090-2449-0/16/$31.00 ©2016 IEEE

way to handle missing values. forecasting, clustering have their own claim when it comes to
Conventional methods are method that emerged on the treating missing values by replacing it with mean or mode, or
early stage of missing values handling study and relatively a even delete the data that containing missing vlaues on their
simple yet risky methods, said to be risky because it is preprocessing step.
solving a problem but creating another problem. To This paper will provide a review of several studies related
overcome the additional problems and flaws caused by the to missing values handling methods or techniques especially
conventional methods, the more modern way to handle on time series data. This paper consist of four sections. The
missing values called imputation is introduced. The further first section contain explanation about related things about
explanation of those missing values handling techniques wil data and missing values. Afterwards, the second section
be provided in the next section. contain some explanation about missing values handling
Table I presents the example of missing values existence methods from the conventional one to the modern one. The
on a dataset where the “?” mark represents the missing third section will contain discussion about the method that
values of the dataset. been explained in the previous section. And last section will
conclude this paper including plausible option of estimation
method.
TABLE I. DATASETS CONTAIN MISSING VALUES
II. MATERIAL AND TECHNIQUES
V1 V2 V3 V4 V5
Y1 0.11 ? 0.42 0.25 0.12 Missing values is not a new topic of studies in data mining
Y2 0.12 0.23 0.25 0.12 0.53 or data analysis since there are so many methods, approach,
Y3 0.13 0.25 ? 0.27 0.26 and techniques are proposed by the previous researchers,
Y4 0.14 0.26 ? 0.23 0.15 from the simplest one to the complicated one and their own
Y5 0.14 0.62 0.62 ? 0.37
advantage and drawbacks. Some basic methods are proposed
Y6 0.16 ? 0.36 ? 0.42
in [5], such as ignoring, deleting, zero or mean or mode
estimation methods. These methods above have the simplicity
Missing values can be distinguished by the two main
but only effective for low percentage of missingness, but in
pattern of missingness as described in Table II. The first
bigger percentage of missingness, the result will affect the
pattern is, the data are missing monotone if we can observe a
result of the analysis and even biased result [6]. The main
pattern among the missing values, it may be necessary to
disadvantage of discarding incomplete observation is the loss
reorder or rearrange variables and/or individuals. The second
of efficiency and biased estimation result especially when the
pattern is the data that are missing arbritarily if the data are
data missingness is systematic [6]. The quality of data mining
can not be reordered or rearranged to obtain the clear pattern
or analysis is influenced by the quality of the data. Therefore,
of the missingness [10].
the data that contain missing values should be estimated to
provide complete case of data to get expected result from the
TABLE II. MISSING VALUES PATTERN [10]
data [6]. Nowadays, there are many studies conducted and
Monotone Arbritary resulting some methods or approaches to solve missing values
x1 x2 x3 x4 x1 x2 x3 x4 problem. The next section will provide an explanation about the
X X X X X X X X methods and approaches to estimate missing values.
X X X X - X X -
X X X - X X X - A. Conventional Method
X X - - - X X X
X - - - X X - X
Many methods have been invented to deal with missing
values, several of them are really simple methods. Some of the
methods using statistical principle as their base.
Table II describes missingness pattern of a dataset by its
dependencies of the data structure. The existing data 1) Ignoring
represented by “X” and the missing values represented by “-”. The first way to deal with missing values as pointed in
[6] is ignoring the missing value, which is the simplest way to
Since the missing values handling will be treated in the
preprocessing step, it is important to identify what we will do deal missing values. The missing values is completely ignored
as the analysis is carry on. Though it is a simple method yet
to the data that contain missing values. The percentage of
very risky if the missingness percentage of the data are big
missingness will affect the method we use to solve it. Paper
that focusing on missing values handling have their own enough to disrupt the result of the analysis [5].
percentage of missing values to test their methods or models. 2) Deletion
Generally, the papers uses missing values percentage of 5%- As what it says, deletion method is simply delete the
50% data [3,11]. But some certain papers uses a bit extreme missing variable, or the instance of observation data to
percentage that ranged from 1%-80% data that contain continue the analysis or data mining process [5,6].
missing values [4,12,14], and in some other paper, it starts 3) Mean/mode Imputation
from 10%-50% [2]. So, the significancy of the missing values While ignoring and deletion did not give a good result
percentage that are need to be treated by specific method is of an analysis or data mining process caused by missing values
not clear enough. Several studies of classification, (Ignoring) and less data to be proceed (Deletion), Mean/Mode
Imputation method comes as a solution to give a better result, it because the newly imputed data are the same with the mean of
solves the missing values problem and the number of data that the observed data.
expected to be proceed is remain to be the same number. But To overcome the bias problem caused by the statistical
the drawback of this method is the bias caused by so many method result of missing values estimation mentioned above,
values on the data are have a similar values [5,6]. several methods have been proposed in the previous studies
that concerning in missing values estimation. The method is
B. Imputation Procedures adapted from the other field of research to be able to solve
Imputation procedures can produce a complete sample of problem in this research field. The latest technique than the
data, but ignores the consequences of the imputation methods statistical one is soft computing techniques that also including
side effect. Imputation or estimation is a procedures that method which is need learning phase to be able to estimate
handling missing values problem by replacing each of the missing values.
missing values by some certain values (the values source is 1) Autoregressive
different for each methods). From the principle of statistical Sridevi et al. [11] proposed a method that focuses on
field, several methods have been proposed, such as Hot and autoregressive-model-based (ARLSimpute) to estimate missing
Cold Deck Imputation, Mean and Mode Imputation, and values. This method is claimed to be able to handle missing
Multiple Imputation. In this section we will give a brief values where a particular time point contains many missing
explanation about these methods. values or where the entire point time is missing. The
1. Hot and Cold Deck Imputation mechanism of the method are, find the K-similar data from the
In hot deck the imputation is done through replacing dataset, then calculate the AR Coefficients for the K-similar
the missing values X by matching the values from diffferent data and last, estimating the missing values. If the result
observed data Y with similar variables X and take the values. coverged then the result will evaluated using Normalized Root
The cold deck is in contrast from hot deck method, the Mean Square Error (NRMSE), if not coverged, the step
imputation is done through replacing the values by the external rollback to find K-similar data. This method then compared
source, like the values from the previous observation dataset with KNNimpute method by measuring the NRMSE result.
with the same domain of observation. These types of method is The NRMSE result of the ARLSimpute and KNN Impute
simple, but it has a bad effect when the dataset have large shown that ARLSimpute performed better than KNN Impute
amount of dataset, moreover when the assumption of missing generally. However, ARLSimpute performing better in lower
data are MAR [14]. order because the error rate is increasing as the order increased.
2. Mean Imputation 2) Genetic Algorithm Optimization Based
One of the easiest way to impute or estimate missing Genetic algorithm (GA) have been widely applied as a
values to get a complete sample is replacing each of the global search in various domains, data mining problems and
missing values with the mean of the observed data for that optimization problems are among the domain. The main
variable, or known as unconditional mean imputation. Aside concerns of choosing GA as a solution for the problem are: (1)
of mean imputation, there is median and mode that can used GAs explore large search spaces while exploit optimal
for replacing missing values [14]. solution; (2) the GAs paradigm is scalable and can be
Kevin strike et al. [15] did an evaluation of three effectively parallelized; (3) GA is relatively easy to implement
missing data handling techniques: listwise deletion, mean and adapt for different domains [16].
imputation, hot deck imptutation for the certain purpose. Lobato et al. [16] introduced a multi-objective genetic
Listwise deletion (one of deletion method) can handle missing algorithm method based on NSGA-II called Multi-objective
values for below 15%, however its accuracy is different to the genetic algorithm imputation (MOGAImp). In many problems,
increasing number of missingness percentage. Euclidean GA operators is very simple and have limited impact on
distance and z-score standarization are employed by hot deck processing time. However, the fitness calculation may demand
method and proves to provide a consistent and accurate high computation time. To solve the problem, paralell
compared to the other methods. But the three methods are less approach also proposed to reduce processing time without
effective for non-ignorable missing values compared to interfering with other algorithm properties. Therefore, the
MCAR and MAR [14]. fitness processing will not take long time to done since each
3. Multiple Imputation individual is assigned to one thread and done at the same time.
Other statistical method of missing values imputation Tang et al. [12] proposed GA as an optimization method
is multiple imputation which proposed by rubin [5]. In this for the fuzzy c-means. FCM as a clustering method needs to be
method missing values imputed n-times to represent the optimized to get the best result from the clustering process. GA
uncertainty of possibile values that are to be imputed. The n- was choosen as the optimization method because of its
times values then analyzed to get a single combined estimates advantages above. GA used on the study to make sure the
[14]. optimum solution (minimum errors) is obtained, the the
The result that produced by statistical method is somewhat missing values is estimated.
questionable because just effective in small amount percentage Aydilek et al. [6] introduced a hybrid method that
of the missing values. For example for mean imputation, the combine FCM, support vector regression (SVR) and GA.
data that is imputed by this method will suffer from high bias Similar to what Tang et al.[12] proposed, but the estimation
mechanism is done through machine learning as SVR is used. values and known values sample time, the weight increases as
The purpose of SVR and GA usage is to minimize the error the sample time distance of the data that is missing is closer to
after the estimation of FCM is done. The mechanism of the the known value and vice versa. IDW then evaluated by
method is as follows : (1) SVR algorithm is trained using the compare it with mean imputation and see the RMSE score
complete dataset rows; (2) the incomplete dataset rows result, the performance shown that IDW outperformed mean
estimated using FCM, and compare the FCM output and the imputation by big margin.
SVR output vector; (3) GA works on the minimizing the 5) Maximum Likehood
difference between FCM and SVR output and obtained Banbura et al. [19] conducted a study of missing values
optimized c and m parameter; (4) estimate the missing values imputation on arbritary pattern of missingness data. The
using optimized parameters using FCM. proposed method is maximum likehood (EM) based method as
Azadeh et al. [17] conducted a study to test some a general method of missing values imputation method. The
estimation methods to be applied in randomized complete steps on the proposed method are fill in the missing data in the
block design table. One of the method is GA, GA run over 30 like of expectation step, and re-optimize the expectation step
times of iteration to make a complete table and claimed to be in the maximization step. On certain regularity condition the
better than other method that are compared in the study. The EM algorithm converges toward the local optimum. The
correlation between the estimated data and the actual data advantage of the maximum likehood is the method allow us to
emerged GA and ANN as the best method and been impose the restrictions on the parameter in a relatively
encouraged to uses GA approach for setimating missing values straightforward manner. This methodology is well performed
in RCBD tables. on the large fraction of missing data and small sample.
3) Support Vector Machine 6) Fuzzy-rough set
Wu et al. [3] proposed a method to handling missing Amiri et al. [1] intoduced fuzzy-rough methods to
values namely least square support vector machine, the basic impute missing values on time series data. The main reason to
idea is “Local Time Index” (LTI) which ignores the missing use this method is that it give an exellent framework to deal
values and adopts temporal information in addition to put in with uncertainty that is often occuring in imputation problem.
training patterns. Any missing values methods that basicaly It simply calculate the fuzzy similarities of instances and can
manipulating the existing time-series data may cause another work with noise and can deal with missing data. The method of
problem. The LTI idea is ignores the missing values yet it want fuzzy-rough used as lower and upper approximators, and the
to keep the continuity of the series itself. The solution is prediction work is given to KNN algorithm. The algorithm is
providing temporal information to replace the missing values, work as follows: for every instances of the dataset “a” that
the temporal information of a pattern somehow can be obtained containing at least at least one missing values for each feature
through some forecasting model. That hipothesis is the basis of “b”, the algorithm find its KNN and puts them in set “X”. The
this method. At the comparison section, this method is approximation done by using a’s nearest neighbor. This
compared to several conventional method such as, hot-decking, method outperformed other variant in the same basis,ordered
mean, and Sridevi et al. [11] proposed method’s AR. This weighted average-based fuzzy-rough sets with nearest neighbor
method is shows a better result than those methods. (OWANNI) and vaguely quantified nearest neighbor
4) Interpolation imputation (VQNNI) in missing values estimation domain.
Shao et al. [18] conducted a study to handle missing 7) Similarity Measure
values by introducing an interpolation method based on Sitaram et al. [4] introducted a method of time series
NURBS geometry modeling. The study is discussed about missing values estimation based on Mahalanobis distance that
interpolation method using window interpolation adjustment is known as a method for handwritten character recognition. It
based on time series features in assumption that the most time is also used widely in data mining domain such as
series of the system have certain cyclical characteristic, which classification and clustering techniques. The method developed
reflects on seasonal volatility. The window interpolation in two phases. In the first phase, the candidate with missing
adjustment is based on such periodic correlation sequence. This values and the query time series is assumed to be exact
method then evaluated using two evaluation index NRMSE matched. The result is formed as best match and exact match
(normalized root mean square error) and COEF (Pearson’s for each algorithm. The mahalanobis shown the same results as
product-moment correlation coefficient) and the result exact match with a good number missing values that implying
outperformed Sbspline as comparison. The Snurbs the reasonable way to extend the algorithm for missing values.
interpolation combined with window adjustment is superior In the second phase, extend the algorithm to also acomodate
than traditional interpolation method. The Snurbs introduction warped time series data and proposed and algorithm that
on the interpolation give a flexibility and provide unified handles warped data using the same way as dynamic time
mathematica expression for time series, and the window warping (DTW) did. The result from this methods is a
adjustment improves the initial interpolation result effectively. complete dataset without performing any imputation in the
Sree dhevi [14] proposed an interpolation method candidate to handle missing values. It is just simply replace the
namely inverse distance weigted(IDW). By this method, the missing series by the one that best matched.
missing values can be predicted using the measured data, the
influence of the distance is affected by the distance of missing
 III. DISCUSSION apparently better that the previous approach, because the
Estimation techniques is required to handle dataset that have parameter is can be vary depend on the method itself. Sridevi
non-ignorable missing data and large amount of missing et al. [11] presented probably the most basic and simple
values. Estimation techniques that have been used by previous estimation method that are based on autoregressive model.
researchers shown good performance to the method that are The method is claimed to be able to solve missing values at
compared to. From the literature review, we can summarize the several time points or columns of the dataset. To prove the
papers based on method, and dataset as described in Table III. method is doing well, it is compared to KNN based estimation
method by the score of NRMSE. The method is better of KNN
TABLE III. LIST OF TIME SERIES MISSING VALUES PAPERS based imputation method in terms of handling such type of
missing values. From the papers above, the most used method
Estimation Techniques
Paper year Method dataset
by previous researchers in interpolation and genetic algorithm
Sridevi et al.[11] 2011 Auroregressive- Stock combination method [6,12,14,16,17,18]. The main reason to
model-based UK statistic choose those methods is their easy to use mechanism and the
missing values Sales flexibility of manipulating the step. The other reason is mainly
estimaton method Weather
because these method is can be easily implemented on various
Aydilek et al.[6] 2013 Optimized fuzzy c- Glass domain.
means with support Haberman From the methods explained, genetic algorithm and the
vector regression Iris combination that include GA [6,12,16,17] and interpolation
and genetic Musk1 methods [14,18] are mostly used to solve missing values
algorithm Wine
Yeast problems. But the most frequently used method can not
Azadeh et al.[17] 2013 Genetic algorithm - guarantee that those are the best method for general missing
Sree dhevi [14] 2014 Inverse distance Istanbul stock values problems, since some problem in some certain dataset
weighted exchange have its own characteristic.
Banbura et al.[19] 2014 Maximum likehood -
Shao et al.[18] 2014 Combination of -
In terms of flexibility, genetic algorithm [6,12,16,17], fuzzy
Snurbs with window c-means [12,16], and autoregressive [11] are considered as the
interpolation best among the menthods explained above. Beside those
adjusment methods can be manipulated according to the needs of the
Wu et al.[3] 2015 Least squares Sine function user, those method can combined with another complementary
support vector Sinc function
machine Poland method.
electricity load
Sunspot IV. CONCLUSION
Jenkins-box gas
Missing values can be treated in various way depends on
furnace
Tang et al.[12] 2015 Fuzzy c-means and Traffic volume the problem caused by the missing values. The method of
genetic algorithm missing values can affect the result of the analysis or data
method integration mining process. Estimation technique is probably the best
Lobato et al.[16] 2015 Multi-objective Public option of missing values handling, since to accomplish certain
genetic algorithm benchmarking
dataset work the complete dataset is required and some dataset have
Sitaram et al.[4] 2015 Mahalanobis - dependend variable which is imposssible to delete the missing
distance values as it can disrupt the data itself. Although the recent
Amiri et al.[1] 2016 Fuzzy-rough set - methods performance is good, the better methods to cut out the
time complexity or demanding computational resource that
In most paper, the accuracy measurement is done by using required by certain method.
NRMSE or RMSE method because they use comparison as This paper figures missing values handling method from
their way to measure the method’s performance from the other the conventional one to the modern one. Methods such as
method since NRMSE and RMSE score represent the deletion, mean imputation, and hot decking are considered as
accuracy of the method. In terms of simplicity, Sitaram et al. conventional method which can be used for more general
[4] provides the most simple way to complete the dataset by dataset, while estimation techniques is modern method that
just replacing the missing values with a part of another values specifically picked to handle missing values in time series data.
in the closest dataset without any imputation or estimation This overview may give researcher state of the art about
process. It means less mathematic equation. However, the missing values estimation which is performed by other
method is fully based on threshold on matching process, in researcher in various methods. This overview may also provide
case of the threshold can not met, the missing values can not a view of each method for what its advantage and disadvantage
be solved. In several time series data there are anomaly to take into consideration of future research in this field of
conditions or the deviation between periods is high. So, the study.
condition when threshold is can not fulfilled is very possible.
The possible solution is decreasing the threshold, but it may REFERENCES
reduce the quality of the method. Mathematical approach is [1] M. Amiri and R. Jensen, “Missing data imputation using fuzzy-
rough methods,” Neurocomputing, pp. 1–13, 2016.
[2] C. Yozgatligil, S. Aslan, C. Iyigun, and I. Batmaz, “Comparison of
missing value imputation methods in time series: The case of
Turkish meteorological data,” Theor. Appl. Climatol., vol. 112, no.
1–2, pp. 143–167, 2013.
[3] S. Wu, C. Chang, and S. Lee, “Time Series Forecasting with
Missing Values,” 2015 1st Int. Conf. Ind. Networks Intell. Syst., pp.
151–156, 2015.
[4] D. Sitaram, A. Dalwani, A. Narang, M. Das, and P. Auradkar, “A
Measure of Similarity of Time Series Containing Missing Data
Using the Mahalanobis Distance,” Proc. - 2015 2nd IEEE Int. Conf.
Adv. Comput. Commun. Eng. ICACCE 2015, pp. 622–627, 2015.
[5] D. Little, RJA and Rublin, “Statistical Analysis with Missing Data,”
Wiley, New York., p. 381, 1987.
[6] I. B. Aydilek and A. Arslan, “A hybrid method for imputation of
missing values using optimized fuzzy c-means with support vector
regression and a genetic algorithm,” Inf. Sci. (Ny)., vol. 233, pp. 25–
35, 2013.
[7] Y. S. Afrianti, “Imputation Algorithm Based on Copula for
Missing,” pp. 252–257, 2014.
[8] A. Briggs, T. Clark, J. Wolstenholme, and P. Clarke, “Missing...
presumed at random: cost-analysis of incomplete data.,” Health
Econ., vol. 12, no. 5, pp. 377–92, May 2003.
[9] J. Sensch, “Book Reviews,” Br. J. Math. Stat. Psychol., vol. 55, no.
1, pp. 193–196, 2002.
[10] Y. Yuan, “Multiple imputation for missing data: Concepts and new
development,” SAS Inst. Inc, Rockville, MD, pp. 1–13, 2010.
[11] S. Sridevi, S. Rajaram, C. Parthiban, S. SibiArasan, and C.
Swadhikar, “Imputation for the analysis of missing values and
prediction of time series data,” 2011 Int. Conf. Recent Trends Inf.
Technol., pp. 1158–1163, 2011.
[12] J. Tang, G. Zhang, Y. Wang, H. Wang, and F. Liu, “A hybrid
approach to integrate fuzzy C-means based imputation method with
genetic algorithm for missing traffic volume data estimation,”
Transp. Res. Part C Emerg. Technol., vol. 51, pp. 29–40, 2015.
[13] X. Zhu, S. Zhang, Z. Jin, Z. Zhang, and Z. Xu, “Missing value
estimation for mixed-attribute data sets,” IEEE Trans. Knowl. Data
Eng., vol. 23, no. 1, pp. 110–121, 2011.
[14] A. T. Sree Dhevi, “Imputing missing values using Inverse Distance
Weighted Interpolation for time series data,” 6th Int. Conf. Adv.
Comput. ICoAC 2014, pp. 255–259, 2015.
[15] K. Strike, K. El Emam, and N. Madhavji, “Software Cost
Estimation with Incomplete Data,” IEEE Trans. Softw. Eng., vol.
27, no. 10, pp. 890–908, 2001.
[16] F. Lobato, C. Sales, I. Araujo, V. Tadaiesky, L. Dias, L. Ramos, and
A. Santana, “Multi-objective genetic algorithm for missing data
imputation,” Pattern Recognit. Lett., vol. 68, pp. 126–131, 2015.
[17] A. Azadeh, S. M. Asadzadeh, R. Jafari-Marandi, S. Nazari-
Shirkouhi, G. Baharian Khoshkhou, S. Talebi, and A. Naghavi,
“Optimum estimation of missing values in randomized complete
block design by genetic algorithm,” Knowledge-Based Syst., vol. 37,
pp. 37–47, 2013.
[18] C. Shao, “An interpolation method combining Snurbs with window
interpolation adjustment,” 2014 4th IEEE Int. Conf. Inf. Sci.
Technol., pp. 176–179, 2014.
[19] M. Bańbura and M. Modugno, “MAXIMUM LIKELIHOOD
ESTIMATION OF FACTOR MODELS ON DATASETS WITH
ARBITRARY PATTERN OF MISSING DATA,” J. Appl.
Econom., vol. 29, no. 1, pp. 133–160, 2014.

View publication stats