A Comparative Study for the Selection of Machine Learning Algorithms

based on Descriptive Parameters

Chettan Kumar, Martin K

appel, Nicolai Sch

utzenmeier, Philipp Eisenhuth and Stefan Jablonski

Institute for Computer Science, University of Bayreuth, Universit

atsstraße 30, Bayreuth, Germany

Keywords:

Machine Learning, Algorithm Recommendation, Data Analysis, Information System.

Abstract:

In this paper, we present a new cheat sheet based approach to select an adequate machine learning algorithm.

However, we extend existing cheat sheet approaches at two ends. We incorporate two different perspectives

towards the machine learning problem while simultaneously increasing the number of parameters decisively.

For each family of machine learning algorithms (e.g. regression, classiﬁcation, clustering, and association

learning) we identify individual parameters that describe the machine learning problem accurately. We arrange

those parameters in a table and assess known machine learning algorithms in such a table. Our cheat sheet is

implemented as a web application based on the information of the presented tables.

1 INTRODUCTION

The development of algorithms for machine learn-

ing tasks has made large progress over the last years.

Nowadays, there is a very large number of differ-

ent algorithms for data analysis problems. Conse-

quently, often the hardest part of solving a machine

learning problem is ﬁnding the right algorithm for the

job. Obviously, it would be ideal to have vast knowl-

edge about strengths and weaknesses of all these al-

gorithms. However, this is almost not possible due to

the incredible number of different algorithms. Also, it

is good to know from experience when to use which

algorithm. Alternatively, a general approach for ﬁnd-

ing the best ﬁtting algorithm is to run the most known

ones and select the three to four best performing ones

for further testing. Obviously, this is a very time

consuming and tedious work without a deeper under-

standing of the underlying causes.

To overcome the problem of selecting an adequate

machine learning algorithm, several cheat sheets are

provided on the internet. The idea is to ask a do-

main expert some questions about the machine learn-

ing task and then recommend one or more algorithms.

For ease of use these cheat sheets are kept very sim-

ple. That means they only focus on a small set of

parameters (mostly size of data and labelling) and ig-

nore important properties of algorithms like the risk

for overﬁtting or the handling of different types of

noise. With regard to the complexity of machine

learning problems we consider such cheat sheets as

not adequate to cope with a decent selection of a ma-

chine learning algorithms (see Section 3). In addi-

tion, there are many interesting research papers (Kot-

siantis, 2007; Xu and Tian, 2015), which compare

different machine learning algorithms and work out

strengths, weaknesses and parameter based recom-

mendations with respect to speciﬁc machine learning

tasks. Nevertheless, the information about these ma-

chine learning algorithms is spread over many papers

and this is very difﬁcult to grasp.

In this paper, we present a new cheat sheet based

approach to select an adequate machine learning al-

gorithm. Therefore, we extend existing cheat sheet

approaches at two ends. First, we incorporate two

different perspectives towards the machine-learning

problem while simultaneously increasing the num-

ber of parameters decisively. While the ﬁrst perspec-

tive provides an application-oriented view, the second

perspective focusses on a set of technical parameters

about the underlying data, algorithms, and their us-

age. The application-oriented view contributes ex-

perience of solving similar problems by describing

the problem with the relevant parameters and the ap-

plied algorithm. The increased number of parame-

ters describing either the application or the technical

perspective enables a more realistic and accurate se-

lection of a machine learning algorithms since both

problem and solutions are described in much more de-

tail. In the best case scenario, the application perspec-

tive and the technical perspective of the cheat sheet

search and propose algorithms, respectively, and the

408

Kumar, C., Käppel, M., Schützenmeier, N., Eisenhuth, P. and Jablonski, S.

A Comparative Study for the Selection of Machine Learning Algorithms based on Descriptive Parameters.

DOI: 10.5220/0008117404080415

In Proceedings of the 8th International Conference on Data Science, Technology and Applications (DATA 2019), pages 408-415

ISBN: 978-989-758-377-3

intersection of both proposals is not empty. The algo-

rithms lying in the intersection should then depict the

best or a set of best ﬁtting algorithms for the problem

to solve. Otherwise, we get a list of algorithms that

does not need to be further tested. With respect to

knowledge, experience, and available time, it is also

possible to ignore one of the perspectives at the ex-

pense of accuracy.

Our second extension of conventional cheat sheets

is the overall structure of our cheat sheet. For each

family of machine learning algorithms (e.g. regres-

sion, classiﬁcation, clustering, and association learn-

ing) we identify individual parameters that describe

the machine learning problem accurately. We arrange

those parameters in a table and assess known machine

learning algorithms in such a table. Since we expect

that further parameters, which might be very helpful

in assessing the adequacy of a speciﬁc algorithms, can

be added easily as the tables are extensible. Typically,

the labelling of these algorithms in that table is based

on an extensive literature review and on practical ex-

periences made in numerous data analysis projects. It

is worth to mention that it is also possible to simplify

the cheat sheet with respect to the experience of the

user or the available information about the data set by

leaving out some columns (parameters). In contrast to

the hierarchical structure of conventional cheat sheets

(see Section 3) not all parameters have to be evaluated

for a recommendation.

The rest of this paper is structured as follows. Sec-

tion 2 gives an overview of related work. In Section

3 we analyse three different existing cheat sheets. Af-

terwards, in Section 4, we deﬁne relevant parameters

for the analysis of machine learning algorithms. In

Section 5 and Section 6, we present our cheat sheet

and give reasons for the judgment of the labelling of

the used parameters. Sections 7 and 8 conclude the

main results of the paper and give an overview of fur-

ther work.

2 RELATED WORK

Selecting an adequate algorithm for a machine learn-

ing problem is a difﬁcult task due to the vast amount

of available algorithms. Hence, working out strengths

and weaknesses is an essential part of a large number

of research papers. In (Kotsiantis, 2007) several clas-

siﬁcation techniques are evaluated with regard to ac-

curacy, speed, tolerance to missing values, redundant

attributes, dealing with danger of overﬁtting or noise.

As a rating system the authors use a scale from one

to four points, describing best (four points) and worst

performance (one point). In (Arora et al., 2013) and

(Kumbhare and Chobe, 2014) some of the more com-

mon algorithms for association learning (like Apriori,

AIS, SETM and so on) are evaluated and an overview

of strengths and weaknesses are given. In (Ruiand

D. Wunsch, 2005) Xu gives a detailed survey about

clustering algorithms and their application. A theo-

retical background is given and the complexity and

the capability of tackling high dimensional data for

the algorithms is evaluated. In (Xu and Tian, 2015)

the survey is expanded by comparing their capability

for data size, the shape of model and their sensitiv-

ity to noise and outliers. Note that in many research

papers no distinction is made between noise and out-

liers. In (Duwe and Kim, 2017) supervised learning

algorithms are tested on different data sets and their

accuracy and speed is measured by AUC, Precision,

Recall and ACC.

3 ANALYSIS OF EXISTING

CHEAT SHEETS

This section gives an overview on some existing

graphical cheat sheets that should help the user choos-

ing an adequate algorithm for a given machine learn-

ing problem. We criticize the following deﬁciencies

in all these cheat sheets: First, the number of at-

tributes that contribute to the recommendation is not

very high. Second, the properties of the algorithms

under selection, like the handling of missing values

and noise or the risk of overﬁtting, are completely ig-

nored. Besides, the authors made no distinction be-

tween the data-oriented and application-oriented per-

spective and association learning is not covered at

all. They all suffers from the problem that if a spe-

ciﬁc question cannot be answered the user gets stuck

and cannot progress any further. This happens due to

the mostly hierarchical structure. Hence, the cheat-

sheets often cannot give a good recommendation. In

the following part, we discuss cheat sheets from Mi-

crosoft Azure, Scikit-learn, and SAS in depth to dis-

cuss the above-mentioned deﬁciencies in detail.The

cheat sheet from Microsoft-Azure

covers four main

categories of algorithms: regression, classiﬁcation,

clustering, and anomaly detection. It has a hierarchi-

cal structure where the paths leading to the categories

are dependent on the goal of the application. A con-

crete algorithm within a category is then selected by

only one additional decision, which oftentimes is not

very meaningful either, because it only looks at prop-

erties like the accuracy and training time of the algo-

https://docs.microsoft.com/de-de/azure/

machine-learning/studio/algorithm-cheat-sheet

A Comparative Study for the Selection of Machine Learning Algorithms based on Descriptive Parameters

409

Table 1: Classiﬁcation table ([1]=(Ke et al., 2017), [2]=(Chen and Guestrin, 2016), [3]=(Dorogush et al., 2017), [4]=(Freund

and Schapire, 1997), [5]=(Hastie et al., 2009), [6]=(Kotsiantis, 2007), [7]=(Bramer, 2013), [8]=(Aggarwal, 2014)).

Algorithm TOD DOD FN IN RN HO SOTD

Logistic Regression num Y N L

[8] [8] [8] [8]

Naive Bayes cat N Y Y Y

[7] [6] [6] [6] [6]

Artiﬁcial Neural Network num N N N N

[6] [6] [6] [6] [6]

k-Nearest-Neighbour cat N N S

[7] [6] [6] [6]

Support Vector Machine num Y N N

[6] [6] [6] [6]

Decision Tree cat, num Y Y S

[6] [6] [6] [6]

Random Forest cat, num L Y Y Y L

[5] [5] [5] [5] [5] [5]

AdaBoost cat L N N N L

[4] [4] [4] [4] [4] [4]

CatBoost cat L Y Y L

[3] [3] [3] [3] [3]

XGBoost num L Y Y L

[2] [2] [2] [2] [2]

LightGBM cat L Y Y L

[1] [1] [1] [1] [1]

rithms. Additionally, distinct aspects of the problems

are covered by the various decisions, so it is hard to

compare the algorithms among each other. The avail-

able data is almost never part of the decision. The

only considered aspect is the number of features. All

in all, this cheat sheet is not profound enough and

ambiguous. Scikit-learn also offers a graphical cheat

sheet

. It covers four main categories of algorithms:

regression, classiﬁcation, clustering and dimensional-

ity reduction. Here, the resulting recommendation is

based upon a chain of binary decisions that mainly

considers the size and type of the dataset. The cheat

sheet has some dead ends with no alternatives and

only covers a small number of different algorithms.

The last cheat sheet under consideration is offered

by SAS

. Here, the algorithms are ﬁrst grouped by

the learning paradigm (supervised and unsupervised).

Then they are further categorized by clustering and

dimension reduction for unsupervised learning and

classiﬁcation together with regression for supervised

learning. Like with the former discussed two cheat-

sheets, the number of decisions leading to the ﬁnal al-

gorithms is quite small and the quality of the branches

varies in regard of the possible choices and the quality

of the answers.

https://scikit-learn.org/stable/tutorial/machine\

learning\ map

https://blogs.sas.com/content/subconsciousmusings/

2017/04/12/machine-learning-algorithm-use/

4 PARAMETERS

In this section, we describe different parameters that

are useful for the recommendation of appropriate

algorithms and we want to use in our cheat sheet.

On the one hand we have parameters which are

important for every algorithm, e.g. the dimension

of the data set. On the other hand we have certain

approach-based parameters which are only signiﬁcant

for some algorithms, e.g. the shape of a model is only

relevant for clustering algorithms. In the following

we deﬁne and explain all the parameters that are used

throughout the paper.

Type of Data (TOD): This parameter indicates if the

algorithm handles numerical (num), categorical (cat)

or spatial data. Most algorithms do not work with

mixed types. Hence a preprocessing to convert the

data into the required type is necessary.

Dimension of Data Set (DOD): Indicates if the

chosen algorithm can work with a large (L) or a small

(S) number of features in the data set.

Outliers (O): An outlier is a record in the data

set which has at least one value that is very high or

very low compared to the values of all the other data

points. We associate an algorithm with Y if it works

well with a data set with more than 20% outliers.

DATA 2019 - 8th International Conference on Data Science, Technology and Applications

410

Table 2: Clustering table ([1] = (Xu and Tian, 2015), [2]=(Tiruveedhula et al., 2016), [3]=(Ruiand D. Wunsch, 2005).

Algorithm SOD DOD O TOD SOM Based

k-Means L S N num convex Partition

[1] [1],[3] [1] [2] [1] [1]

PAM S S Y cat, num convex Partition

[1] [1] [1] [2] [1] [1]

CLARA L S Y num convex Partition

[1] [1],[3] [1] [2] [1] [1]

CLARANS L S Y num convex Partition

[1] [1],[3] [1] [2] [1] [1]

BIRCH L S Y num convex Hierarchical

[1] [1],[3] [1] [2] [1] [1]

CURE L L Y num, cat arbitrary Hierarchical

[1] [1],[3] [1] [2] [1] [1]

ROCK L L Y cat arbitrary Hierarchical

[1] [1] [1] [2] [1] [1]

Chameleon L L Y num, cat arbitrary Hierarchical

[1] [1] [1] [2] [1] [1]

DBSCAN L S Y num arbitrary Density

[1] [1],[3] [1] [2] [1] [1]

OPTICS L S Y num arbitrary Density

[1] [1] [1] [2] [1] [1]

DENCLUE L L Y num arbitrary Density

[1] [1][3] [1] [2] [1] [1]

STING L S Y spatial arbitrary Grid

[1] [1] [1] [2] [1] [1]

Wave L S Y num arbitrary Grid

[1] [3] [1] [2] [2] [2] [1]

CLIQUE L [1] S [3] [1] Y [1] num [2] arbitrary [1] Grid [1]

SOM S L Y num, cat convex Model

[1] [1] [1] [2] [1] [1]

SLINK L S N num arbitrary Model

[2] [2] [2] [2] [2] [1]

Otherwise we denote it with N.

Feature Noise (FN): Indicates if the chosen al-

gorithm takes measures to work with features that

do not help explaining the target. In other words:

irrelevant or weak features.

Item Noise (IN): Indicates if the chosen algo-

rithm works well with anomalies, like wrong or

missing values, in certain data items.

Record Noise (RN): Indicates if the chosen al-

gorithm works well with records, which do not

follow the form or relation, which most of the records

do.

Handling of Overﬁtting (HO): Indicates if the

chosen algorithm takes measures to avoid overﬁtting.

Size of Training Data (SOTD): Indicates if the

chosen algorithm needs a large (L) or a small (S)

training data set to generate a well-performing model.

As aforementioned this judgment also depends on the

use case.

Size of Data Set (SOD): Indicates if the chosen

algorithm works well with large (L) or small (S) data

sets. If a data set is described as large or small always

depends on the use case, e.g. 10000 image records

would be large whereas the same number of sensor

measurements would be small.

Multicollinearity (MC): Indicates if the chosen

algorithm works well with multicollinearity. Multi-

collinearity is a phenomenon in which two or more

predictor variables in a regression model are highly

correlated, meaning that one can be linearly predicted

from the others with a substantial degree of accuracy.

Based: Describes the basic concepts behind the

clustering. We distingiush between model-, density-,

grid- and hierarchical-based clustering algorithms.

Shape of Model (SOM): Describes the form of

the resulting clusters, e.g. convex or rectangle

clusters.

A Comparative Study for the Selection of Machine Learning Algorithms based on Descriptive Parameters

411

5 RELEVANT MACHINE

LEARNING TECHNIQUES

In this section we present our main result, a table

based cheat sheet for machine learning algorithms.

Therefore, we payed attention on four different tech-

niques: Classiﬁcation, Clustering, Regression and

Association Learning. The following chapters de-

scribe these four techniques and present the tables

with are essential for our developed cheat sheet. An

empty entry means that it is either hard or even im-

possible to ﬁll the cell or we did not ﬁnd a reliable

source to justify our decision.

5.1 Classiﬁcation

Classiﬁcation is an approach of supervised learning

technique for modeling and predicting categorical

variables(Bramer, 2013). The classiﬁcation includes

two phases. The ﬁrst phase is a learning process phase

in which the training data is analyzed, then the rules

and patterns are created. The second phase is the

real classiﬁcation. The in phase 1 trained model is

used to predict the class for the new, not known data

points. Our cheat sheet approach for the Classiﬁca-

tion is shown in table 1.

5.2 Clustering

Clustering is an unsupervised learning technique that

involves the grouping of data points that are in some

properties identical to one another (Bramer, 2013).

It is a common technique for statistical data analy-

sis that used in many different ﬁelds. Basically, clus-

tering algorithms are used to classify each point in

the data set into a speciﬁc group based on the proper-

ties or features. In data analysis, clustering is used to

gain some useful information when we do not have la-

beled data like supervised learning and we are aiming

at ﬁnding out, in which group data points fall when

we apply a clustering algorithm. Our cheat sheet ap-

proach for the Classiﬁcation is shown in table 2.

5.3 Regression

Regression analysis is an important approach for

modeling and analyzing the relationship between a

dependent and one or more independent variables (of-

ten also called explanatory variable). The case of

one independent variable is called simple linear re-

gression, meanwhile for more than one independent

variable the approach is called multiple linear regres-

sion. In both cases the resulting model is called linear

model, because the dependent variable is expressed

as a linear combination of the regression coefﬁcients.

It is used for forecasting, time series modeling and

ﬁnding the causal effect relationship between the vari-

ables. Our cheat sheet approach for the Regression is

shown in table 3.

5.4 Association Learning

The main usage of association rule learning is

for knowledge discovery in transactional databases.

There (mostly) categorical data of grouped items (so

called item sets) is analyzed and association rules are

derived. These rules consider the most associated

items from the dataset based on predeﬁned threshold

values for a support and conﬁdence parameter. The

support parameter for a speciﬁc item set reﬂects the

relative amount of item sets in which it occurs com-

pared to the whole dataset. However, the conﬁdence

parameter describes the relative amount of item sets

in which the right-hand-side of a speciﬁc rule occurs

given that the left-hand-side of the rule is already ful-

ﬁlled (Agrawal et al., 1993). Association rules can be

helpful in many different application ﬁelds like basket

data analysis, cross-marketing, catalog design, web

usage mining, intrusion detection, continuous produc-

tion and bioinformatics. In basket data analysis for

example, a supermarket collects information about

the shopping behavior of its customers. By looking

at products which are purchased together frequently,

valuable information can be obtained. This can then

be used to increase the revenue for the market with

the support of a good marketing strategy based on

the discovered knowledge. The AIS (Agrawal et al.,

1993) and SETM (Houtsma and Swami, 1995) algo-

rithms were early developed algorithms for mining

such association rules from a given dataset. With

the introduction of the Apriori algorithm (together

with AprioriTid and the reﬁned version AprioriHy-

brid, which uses the best parts from both the normal

Apriori and AprioriTid), the performance of ﬁnding

association rules was signiﬁcantly improved by the

clever usage of properties for frequent item set gener-

ation (Agrawal and Srikant, 1994). Eclat (Zaki, 2000)

(short for “equivalence class transformation”) further

reduced the time needed for deriving the rules mainly

by the usage of a vertical database schema for cal-

culating support and conﬁdence values and therefore

reducing the needed database scans. The FP-growth

algorithm avoids the costly candidate generation and

uses a special FP-tree (frequent pattern tree) instead

for deriving the association rules. This improves the

runtime even further, because only two database scans

are needed for the tree generation (Han et al., 2000).

Since the presented algorithms are always an im-

DATA 2019 - 8th International Conference on Data Science, Technology and Applications

412

Table 3: Regression table([1]= (Chen and Guestrin, 2016), [2]=(Tibshirani, 1996), [3]=(Montgomery et al., 2015),

[4]=(Breiman, 2001), [5]=(Hastie, 2017), [6]=(Breiman, 1996a), [7]=(Rodriguez and Yao, 2013), [8]=(Breiman, 1996b),

[9]=(Unger et al., 2009), [10]=(Elith et al., 2008), [11]=(Huang et al., 2006), [12]=(Stone, 1985), [13]=(Friedman, 1991),

[14]=(Drucker et al., 1997), [15]=(Griggs, 2013), [16]=(Marsh and Cormier, 2011), [17]=(Winship and Mare, 1984),

[18]=(Murphy, 2012), [19]=(Hinde, 1982), [20]=(Elster et al., 2015), [21]=(N. van Wieringen, 2015), [22]=(Wulu et al.,

2002), [23]=(Zou and Hastie, 2005), [24]=(Koenker and Bassett, 1978), [25]=(Koenker and Hallock, 2001), [26]=(Luo et al.,

2015)).

Algorithm SOD DOD O HO MC

Linear Regression S S N N N

[7] [3] [7] [3],[18] N [3]

Quantile L S Y N

[24],[25] [24],[25] [24] [3]

Bayesian L S Y Y

[20] [18],[20] [18] [20],[18]

LASSO L Y Y Y

[2] [2] [2] [2]

Ridge L L Y Y Y

[18] [18],[21] [3],[18] [3],[18] [3],[21]

Elastic Net L L Y Y Y

[18],[23] [18],[23] [18] [23] [18],[23]

Ordinal L L Y Y

[17] [17] [17] [17]

Poisson L L Y

[19],[22] [19],[22] [19],[22]

SVR S L Y Y

[14] [14] [14] [14]

Spline L Y Y

[15] [16],[15] [16],[15]

MARS L L Y Y

[13] [15],[13] [13] [15],[13]

Additive L L Y

[12] [12] [12]

RF L L Y Y

[4] [4] [4] [4]

Extreme Learning L Y N Y

[11] [11] [26] [11]

Ensemble L Y Y Y

[9],[8] [9],[8] [9],[8] [9]

Boosted L Y Y

[6],[10] [6],[10] [6],[10]

XGBoost L L Y Y

[1] [1] [1] [1]

provement of the previously presented algorithms, it

is hard to make a fair and useful comparison. The

application of association rule mining is strongly de-

termined by the concrete use case. Therefore, the data

already has to be in the right format to apply the asso-

ciation learning algorithms at all. The quality of the

results from all these algorithms is the same (for same

support and conﬁdent thresholds), which is why the

runtime (and maybe the memory usage) is the only

property for comparison. The most promising algo-

rithms considering the runtime needed for ﬁnding as-

sociation rules within a speciﬁc dataset are FP-growth

and Eclat, because they are the most reﬁned algo-

rithms. All the other algorithms are now mainly use-

ful for educational purposes. There are still some al-

gorithms which were not considered here, since they

are not very well known and they provide no improve-

ment performance-wise.

6 IMPLEMENTATION

Our implementation of the cheat sheet is based on the

tables of the previous section. The content of the ta-

bles is stored in a database together with additional in-

formation like the sources, comments and further rec-

ommendations. We implemented our cheat sheet as

a web application

, which has access to the database

for algorithm recommendation. The implementation

Demo will be available at: https://www.ai4.

uni-bayreuth.de/en/research/ToolsAndResources/

A Comparative Study for the Selection of Machine Learning Algorithms based on Descriptive Parameters

413

is divided into four different parts. The main part of

the cheat sheet is the algorithm recommendation tool,

where an adequate algorithm is proposed based on the

problem description and some properties of the data

set. Here the information which is entered by the user

regarding the properties gets transformed into a corre-

sponding query to the underlying database. For later

maintenance and further extensions, we have imple-

mented a graphical user interface that makes it easier

to edit algorithms and their properties. The applica-

tion dialog enables every user to contribute his expe-

rience into the cheat sheet. All deposited information

can be viewed in the comparison section of the appli-

cation. Algorithms can be listed or sorted by different

criteria and can be compared to each other to inform

oneself or for getting a manual algorithm recommen-

dation.

7 FUTURE WORK

Despite of the large number of evaluated algorithms

and parameters, some table entries still remain blank

and need to be evaluated in a future work. More-

over, some ﬁelds of machine learning, like deep learn-

ing or reinforcement learning were not evaluated yet.

The evaluation of different algorithms in this paper is

based on an exhaustive literature review, but an exper-

imental validation is still missing. Beyond the spe-

ciﬁc task of algorithm comparison, there are many

other open problems that will need to be tackled if

we want to make true progress in solving the grand

task of algorithm recommendation: The judgement

of the parameters depend on so many different fac-

tors (e.g. context, the strength of the parameter or the

preprocessing), that often an evaluation with ”yes”

or ”no” is not sufﬁcient or misleading. Hence we

should try to work out conditions for the evaluation

and include them into our cheat sheet. Therefore, the

comment possibility in our web application should be

improved. A further promising research area for al-

gorithm recommendation is the analysis of success-

fully completed machine learning problems. Based

on a concise description of the problem and the solv-

ing strategy we want to transfer this knowledge to

similar machine learning problems. The cheat sheet

should be extended by increasing the number of anal-

ysed projects. For consistent evaluation of the data

set properties, an automatic detection of the important

parameters would also be useful.

8 CONCLUSION

In this paper we have extended existing cheat sheets

at two ends to a new table based cheat sheet. We

increased the number of parameters in comparison

to the existing cheat sheets decisively and added a

knowledge base to include human experience from

other similar machine learning problems. We eval-

uated the most common algorithms of classiﬁcation,

clustering, regression and association learning by an

exhaustive literature review. Therefore, we choose

for each type of machine learning important proper-

ties to evaluate strengths and weaknesses for the al-

gorithms. A simple judgement with “yes” or “no” re-

spectively “large” or “small” simpliﬁes the algorithm

recommendation because the user must only decide

between two different options. For the ease of use and

better expandability and maintenance, we built a web

application. This application provides a simple query

tool for algorithm recommendation and various input

ﬁelds for updating parameters, adding algorithms or

already solved machine learning problems.

REFERENCES

Aggarwal, C. C. (2014). Data Classiﬁcation: Algorithms

and Applications. Chapman & Hall/CRC, 1st edition.

Agrawal, R., Imielinski, T., and Swami, A. (1993). Min-

ing Association Rules Between Sets of Items in Large

Databases, SIGMOD Conference, volume 22, pages

207–.

Agrawal, R. and Srikant, R. (1994). Fast algorithms for

mining association rules in large databases. In Pro-

ceedings of the 20th International Conference on Very

Large Data Bases, VLDB ’94, pages 487–499, San

Francisco, CA, USA. Morgan Kaufmann Publishers

Inc.

Arora, J. B., Bhalla, N., and Rao, S. J. (2013). A review on

association rule mining algorithms.

Bramer, M. (2013). Principles of Data Mining. Springer

Publishing Company, Incorporated, 2nd edition.

Breiman, L. (1996a). Bias, variance , and arcing classiﬁers.

Technical Report 460, Statistics Department, Univer-

sity of California.

Breiman, L. (1996b). Stacked regressions. Machine Learn-

ing, 24(1):49–64.

Breiman, L. (2001). Random forests. Machine Learning,

45(1):5–32.

Chen, T. and Guestrin, C. (2016). Xgboost: A scalable

tree boosting system. In Proceedings of the 22Nd

ACM SIGKDD International Conference on Knowl-

edge Discovery and Data Mining, KDD ’16, pages

785–794, New York, NY, USA. ACM.

Dorogush, A. V., Ershov, V., and Gulin, A. (2017). Cat-

boost: gradient boosting with categorical features sup-

port. CoRR, abs/1810.11363.

DATA 2019 - 8th International Conference on Data Science, Technology and Applications

414

Drucker, H., Burges, C. J. C., Kaufman, L., Smola, A. J.,

and Vapnik, V. (1997). Support vector regression ma-

chines. In Mozer, M. C., Jordan, M. I., and Petsche,

T., editors, Advances in Neural Information Process-

ing Systems 9, pages 155–161. MIT Press.

Duwe, G. and Kim, K. (2017). Out with the old and in

with the new? an empirical comparison of supervised

learning algorithms to predict recidivism. Criminal

Justice Policy Review, 28(6):570–600.

Elith, J., Leathwick, J. R., and Hastie, T. (2008). A working

guide to boosted regression trees. Journal of Animal

Ecology, 77(4):802–813.

Elster, C., Klauenberg, K., Walzel, M., W

ubbeler, G., Har-

ris, P., Cox, M., Matthews, C., Smith, I., Wright, L.,

Allard, A., Fischer, N., Cowen, S., Ellison, S., Wil-

son, P., Pennecchi, F., Kok, G., van der Veen, A., and

Pendrill, L. (2015). A guide to bayesian inference for

regression problems.

Freund, Y. and Schapire, R. E. (1997). A decision-theoretic

generalization of on-line learning and an application

to boosting. J. Comput. Syst. Sci., 55(1):119–139.

Friedman, J. H. (1991). Multivariate adaptive regression

splines. Ann. Statist., 19(1):1–67.

Griggs, W. (2013). Penalized spline regression and its ap-

plications.

Han, J., Pei, J., and Yin, Y. (2000). Mining frequent pat-

terns without candidate generation. In Proceedings

of the 2000 ACM SIGMOD International Conference

on Management of Data, SIGMOD ’00, pages 1–12,

New York, NY, USA. ACM.

Hastie, T. (2017). Generalized additive models. In Fager-

berg, J., Mowery, D. C., and Nelson, R. R., editors,

Statistical models in S, chapter 7, pages 249–307.

Routledge.

Hastie, T., Tibshirani, R., and Friedman, J. (2009). The el-

ements of statistical learning: data mining, inference

and prediction. Springer, 2 edition.

Hinde, J. (1982). Compound poisson regression models. In

Gilchrist, R., editor, GLIM 82: Proceedings of the In-

ternational Conference on Generalised Linear Mod-

els, pages 109–121, New York, NY. Springer New

York.

Houtsma, M. and Swami, A. (1995). Set-oriented mining

for association rules in relational databases. In Pro-

ceedings of the Eleventh International Conference on

Data Engineering, pages 25–33.

Huang, G.-B., Zhu, Q.-Y., and Siew, C.-K. (2006). Extreme

learning machine: Theory and applications. Neuro-

computing, 70(1):489 – 501. Neural Networks.

Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W.,

Ye, Q., and Liu, T.-Y. (2017). Lightgbm: A highly

efﬁcient gradient boosting decision tree. In Guyon, I.,

Luxburg, U. V., Bengio, S., Wallach, H., Fergus, R.,

Vishwanathan, S., and Garnett, R., editors, Advances

in Neural Information Processing Systems 30, pages

3146–3154. Curran Associates, Inc.

Koenker, R. and Hallock, K. (2001). Quantile regression.

Journal of Economic Perspectives, 15(4):143–156.

Koenker, R. W. and Bassett, G. (1978). Regression quan-

tiles. Econometrica, 46(1):33–50.

Kotsiantis, S. B. (2007). Supervised machine learning:

A review of classiﬁcation techniques. In Proceed-

ings of the 2007 Conference on Emerging Artiﬁcial

Intelligence Applications in Computer Engineering:

Real Word AI Systems with Applications in eHealth,

HCI, Information Retrieval and Pervasive Technolo-

gies, pages 3–24. IOS Press.

Kumbhare, T. A. and Chobe, S. V. (2014). An overview of

association rule mining algorithms.

Luo, X., Chang, X., and Ban, X. (2015). Extreme learn-

ing machine for regression and classiﬁcation using l1-

norm and l2-norm. In Cao, J., Mao, K., Cambria,

E., Man, Z., and Toh, K.-A., editors, Proceedings of

ELM-2014 Volume 1, pages 293–300, Cham. Springer

International Publishing.

Marsh, L. and Cormier, D. R. (2011). Spline regres-

sion models. Journal of Applied Business Research

(JABR), 19.

Montgomery, D. C., Peck, E. A., and Vining, G. G. (2015).

Introduction to Linear Regression Analysis. John Wi-

ley & Sons, New York.

Murphy, K. P. (2012). Machine Learning - A Probabilistic

Perspective. MIT Press, Cambridge.

N. van Wieringen, W. (2015). Lecture notes on ridge re-

gression.

Rodriguez, R. N. and Yao, Y. (2013). Five things you should

know about quantile regression. In In Proceedings of

the SAS Global Forum 2017 Conference.

Ruiand D. Wunsch, X. (2005). Survey of clustering al-

gorithms. IEEE Transactions on Neural Networks,

16(3):645–678.

Stone, C. J. (1985). Additive regression and other nonpara-

metric models. Ann. Statist., 13(2):689–705.

Tibshirani, R. (1996). Regression shrinkage and selection

via the lasso. Journal of the Royal Statistical Society

(Series B), 58:267–288.

Tiruveedhula, S., Sheela Rani, C., and Narayana, V. (2016).

A survey on clustering techniques for big data mining.

Indian Journal of Science and Technology, 9:1–12.

Unger, D. A., van den Dool, H., O’Lenic, E., and Collins,

D. (2009). Ensemble regression. Monthly Weather

Review, 137(7):2365–2379.

Winship, C. and Mare, R. D. (1984). Regression models

with ordinal variables. American Sociological Review.

Wulu, J., Singh, K., Famoye, F., Thomas, T., and McGwin,

G. (2002). Regression analysis of count data. Journal

of Indian Society of Agricultural Statistics, 55:220–

231.

Xu, D. and Tian, Y. (2015). A comprehensive survey

of clustering algorithms. Annals of Data Science,

2(2):165–193.

Zaki, M. J. (2000). Scalable algorithms for association min-

ing. IEEE Transactions on Knowledge and Data En-

gineering, 12(3):372–390.

Zou, H. and Hastie, T. (2005). Regularization and variable

selection via the elastic net. Journal of the Royal Sta-

tistical Society, Series B, 67:301–320.

A Comparative Study for the Selection of Machine Learning Algorithms based on Descriptive Parameters

415