Synergies of Text Mining and Multiple Attribute Decision Making: A Criteria Selection and Weighting System in a Prospective MADM Outline

In this study, a new way of criteria selection and a weighting system will be presented in a multi-disciplinary framework. Weighting criteria in Multi-Attribute Decision Making (MADM) has been developing as the most attractive section in the field. Although many ideas have been developed during the last decades, there is no such great diversity that can be mentioned in the literature. This study is looking from outside the box and is presenting something totally new by using big data and text mining in a Prospective MADM outline. PMADM is a hybrid interconnected concept between the Futures Studies and MADM fields. Text mining, which is known as a useful tool in Futures Studies, is applied to create a widespread pilot system for weighting and criteria selection in the PMADM outline. Latent Semantic Analysis (LSA), as an influential method inside the general concept of text mining, is applied to show how a data warehouse’s output, which in this case is Scopus, can reach the final criteria selection and weighting of the criteria.


Definition of the Current Study in the MADM Outline
Multiple Criteria Decision Making (MCDM) is a multidisciplinary area and field that is working actively in interdisciplinary atmospheres of such fields like management science, operations research and decision science [1,2]. MCDM has two separated parts, which are Multi-Attribute Decision Making (MADM) and Multi-Objective Decision Making (MODM), and it can be described shortly as follows: MADM problems can be considered as discrete problems and MODM problems as continuous problems [3,4]. The MADM structure is linked to the theory of Rational Choice, which is acting rationally with given information, constraints and conditions. Decisions can be made based on alternatives, criteria and the relative importance of them. On the other hand, the MODM framework is designed for non-predetermined alternatives, in which decision makers are involved in to find one of a set of appropriate answers for their models. Generally, the number of alternatives for a MODM problem is infinite [5,6].
At the present time, MADM models and methods are reflected and applied for decision-making problems in different majors and fields, which is not limited to the any special area or structure [7][8][9][10][11][12][13].
In the next section, all MADM methods will be introduced in Table 1. The main point about all common MADM methods is they can be categorized in certain sections that are usually predictable. It means all new contributions could be classified in common sections, which can be comparable as well. The four main contributions as categories are as follows: Table 1. Primary model of Prospective Multi-Attribute Decision Making (PMADM) based on limiters and boosters [44].  The main idea of the current study is to present a new way to weigh criteria based on something more scientific. Although using experts' opinions have had a great position in the decision-making history, results would not necessarily be accurate and robust [18]. Furthermore, one step backward is the criteria selection strategies usually conducted by researchers based on limited previous studies.
Criteria selection, itself, can be an essential part of defining an MADM model and MADM problem in reality. Accordingly, this study can be classified as in Section 1, and even as something newer that has not yet been added as a category of innovation in the MADM field. Text mining and its analysis can be a really powerful tool for finding the most critical criteria based on the entire data base, and then weighting the criteria based on the majority of existing reports and analysis of similar studies.

Definition of the Current Study in the Prospective MADM Outline
PMADM is a new approach and model for decision-making about the future in practice. Since introducing this new approach, a new sub-branch has been imaginable in the MADM framework, which can be developed more and more in reality. As can be analyzed from the literature of the MADM framework, some studies have been working since 1988 about the MADM structure and its framework in decision-making for future matters and topics [45][46][47][48][49][50][51][52][53][54][55][56]. Time (time period) consideration has been developed in decision-making problems and Dynamic MADM with different definitions could be considered as the last contribution regarding time consideration in MADM models.
The MADM and classic methods used to consider a decision in a stable and fixed state that could not be flexibly measured. Dynamic MADM (DMADM) has developed since around one decade ago but could not meet and support all necessities, needs and requirements. By developing "Futures Studies" and "Foresight" perspectives, imaginations and thoughts about the decision process about the future have changed. Classic decision-making structures could not meet such ideas like explorative and descriptive perspectives, so new paradigms and ideas have shaped since then. PMADM was introduced to cover and support all new aspects of needs and necessities of decision making about the future with a flexible idea.
PMADM as a new sub-branch in the MADM field, which is also a multidisciplinary area, and it can be considered as an approach in "Futures Studies" as well. PMADM is not limited to the classic dimensions of MADM and it can be developed in a really new space. Due to needs, new items can be added to the classic model and make that more applicable with more reliable and accurate outputs and decisions. In the first step, Limiters (L) and Boosters (B) are presented [44]. Limiters and Boosters as new items that will be considered in cases in which different scenarios can happen with different possibilities. Mostly it considers alternatives that will have a different quality in different scenarios. Hashemkhani Zolfani et al. [57,58] discussed the importance of considering the future in MADM models. Another new model has discussed about MADM framework and future scenarios in different states and situations [59].
PMADM has this potential to be developed in both concept and for introducing new methods that have the same framework. New items and rules can be added and considered in evaluating criteria and their weights, alternatives and the general concept. New methods with the basic structure of PMADM also can be developed for application in the future in a better way in real-world cases. Here, the main point is methods can be developed the same, original PMADM structure.
All the latest contributions of Prospective MADM are based on new items: • Limiters/Boosters: Hashemkhani Zolfani et al. [57] released the first contribution and definition of the PMADM outline, which can be explained as Limiters/Boosters. Limiters/Boosters can have the role of pay-offs of future scenarios for the evaluated alternatives in their positions. This fact can be demonstrated by examining where they are located in the structure of a classic MADM structure; for example, as shown in Table 1. Limiters and Boosters can be outputs of some future scenarios or just some future possibilities and can have a direct influence on the alternatives' analysis and expectations. •

Multi-Aspect Criterion
Multi-Aspect Criterion is a new item in the classic structure of MADM in the PMADM area. It contains two main shapes: "Hybrid criteria as a new criterion" and "a lately defined concept for the other criteria as a criterion". The importance of time will be showed with this new item to control the definitions during the period of time. In future definitions, criteria can be mixed or developed in different aspects and approaches. It is really important to have an explicit definition about a certain time in the future while the decision-making process is happening [60].

•
Supportive-backup criteria "Supportive-backup criteria" is another additive item to the PMADM outline. While different future scenarios are considered, this new item can be really useful. It shapes all future decision-making matrices into one matrix that decision-makers can shape to whatever they want and make their decisions better and more effective [61]. For instance, an example is illustrated in Table 2. "Supportive-backup criteria" gives great possibility to the researchers to consider a set of different future-possible scenarios for their calculations and evaluations. Decision-makers can have a back-up system for possible ways of managing and leading probable future scenarios in their decision models in advance. Table 2. Position of the "Supportive-backup criteria" [61].  Applying CLA as a qualitative "Futures Studies" field can give a great opportunity to the researchers of the MCDM field to evaluate many things, including analysis of experts who are going to be invited as a part of panel teams. This study showed how experts can be finally selected for cooperation in a study [62].

Research Gap and Case Study: "Machine Tool Selection"
Text mining has been accepted as a powerful and useful tool in foresight exercises [63]. Saritas & Burmaoglu [64] presented Text Mining as a method or tool in the field of Futures Studies, which includes Foresight as well. Prospective MADM is the output and interconnection of two multidisciplinary fields, namely Futures Studies and Multiple Criteria Decision Making. As a matter of fact, this research has a connection between three fields of study and can develop more possibilities for doing better studies in the future.
Making a critical decision for the future is so challenging and all the procedures can be really vital. Similar to the Multi-Aspect criterion, the procedure of defining criteria is a big challenge; therefore, reliability of selecting the most important criteria should be the core of an MADM problem and challenge. If researchers cannot define the best set of criteria, a logical, useful output would be out of reach.
Text Mining can help researchers to use big data to find the most important criteria and relative importance of each of them. The common way of defining a set of criteria in the MADM field is only a limited study field, and in a related literature review, a maximum of some interviews with limited accessible experts are available. Indeed, in numerous studies there is no need to use some older MADM methods (Category 1) for weighting criteria if we are working on future-based decisions in a big level of the study.
Machine tool selection has always been an important issue for decision-makers in different industries in order to make the most efficient and effective decisions. Over the past decades, many researchers studied, with various methodologies, which MCDM is the most well-known methodology that has been used several times. In all the past articles, the criteria and methodologies were selected based on the author's opinion; however, in this investigation we would like to implement a new way in order to classify the most relevant studies and the most significant themes. More specifically, we try to answer the following question: What are the main criteria and topics of current machine tool selection research?
To answer this question, the literature items or literature positions were analyzed by Latent Semantic Analysis (LSA), which revealed five important criteria on machine tool selection: size and precision, cost and serviceability, flexibility, productivity as well as technical features and safety. This study contributes to the machine tool selection literature by providing a comprehensive review of current machine tool selection studies and recognizing its primary research topics, which provides guideline for future studies. This new approach is a unique way to gather data and determine criteria by using text mining based on previous studies and this study would be able to serve as a research map for future MCDM articles.

Data Collection
In order to attain all the related articles in the machine tool selection research landscape and identify the research area, we first searched for the "machine selection" and "MCDM" phrases-for all peer-reviewed academic publications-in their titles, abstract or author-supplied keywords in the SCOPUS database. This database presents the largest abstract and citation of peer-reviewed literature in scientific, technical, medical and social sciences. This process resulted in 107 articles, which means 107 publications exactly used the terms "machine selection" and "MCDM" in their titles, abstract or keywords. Then, the authors read all 107 publications in order to find out the most relevant studies. After reading all 107 abstracts, 28 publications were chosen to be considered as machine tool selection by the MCDM technique research. These 28 articles explain how to select the best machine tool by different MCDM techniques, in which most of the authors used hybrid techniques in order to find the best choice. The abstracts of these articles turned into a new raw data set for analyzing in Latent Semantic Analysis.

Data Analysis
With help of RapidMiner [65] a text mining approach was implemented that is a part of data mining tools. Distinct techniques form the text mining structure, such as natural language processing (NLP), machine learning, information extraction, information retrieval and statistics. This idea derives from "the machine supported analysis of text" [66]. Since this era is famous for data science, there have been many studies that investigated the use of text mining in the literature [67,68]. Latent Semantic Analysis is used in this study to extract the most significant and relevant criteria from the previous studies, which was implemented in different contributions [69,70]. In comparison to other text mining techniques that are only able to analyze textual data and count the occurrences of particular words, LSA can extract the contextual-usage meaning of words and estimates of similarities among words with the information at the semantic level [71]. Thus, the intuitive application of LSA has been growing in different sorts of text mining classifications, containing library indexing, search engine and natural language processing, and so on [72].
This study follows the text mining procedures that was used in previous studies [67,68,71,72]. The following steps explain all processes of the LSA, from the pre-processing term reduction and term frequency matrix transformation to the singular value decomposition.
First, we consider all abstracts as input data in this text mining technique; however, it did not work out very well because usually abstracts do not contain criteria and mostly discuss methodologies and their achievements. Therefore, we read all 28 publications in order to pull out the specific part that has the criteria in its context and then consolidated it in a spreadsheet, finally loading it into RapidMiner. It might form a doubt why we did not consider all parts of the publication as our input data: all the sections of an academic publication together may contain many different contexts and ideas, such as methodologies, literature, mathematic formulas, etc. Thus, we thought it would not a good decision to consider each article completely.

Pre-Processing and Term Reduction
In the first step, every record (the specific part pulled out of the publications) in the dataset is defined as a unique document. This function lets authors trace the results of the LSA back to a specific article to find out which one is of more significance. Secondly, in RapidMiner, the data were imported and called to the procedure by the retrieve operator. Each record was converted into a document object before it could be analyzed. Next, all the words were recognized as tokens and each token was diagnosed by space or a non-letter separator. Then, all tokens were transformed into the lower case, because it is essential to integrate all tokens in a unit format. For instance, "Machine Tool Selection" was considered to "machine tool selection". After that, the "stopwords English" operator removed all stopwords, such as "the", "is", "and", "a", "an", etc., in the English language. The presence of these stopwords do not make any valuable meaning and increases dimensionality. Afterwards, all the tokens that were less than two letters or more than twenty-five letters were removed because in none of both situations the tokens do not make sense. Then, the "stem porter" operator, which is one of the stemming techniques, was applied in order to decrease the number of words that has the same root. For example, "contribute", "contribution", "contributed" and "contributions" were considered as a single token, the "contribut". With the stemming process, plenty of similar words were decreased and this issue helps the dimensionality from a further increment. The result of these processes was concluded at 224 tokens. In this step, we realized that there are some common academic words that are used in academic publications. In order to eliminate these common academic tokens, we searched and selected an academic phrase bank [73]. This academic phrase bank specifically discusses phrases that are used in academic publications. In this step, we considered this book as a discrete input in another procedure and implemented the previous steps containing the tokenization, transforming to lower case, filtering English stopwords, filtering tokens by length and stemming, from where we took the result of 1907 tokens to a new dictionary in order to eliminate all the common academic words in the main process. After this reduction, the number of tokens decreased to 101.

Term Frequency Matrix Transformation
In this study, the technique of calculating the relatively rare weighting called Term Frequency-Inverse Document Frequency (TF-IDF) is used. The TF-IDF technique is a new approach of term frequency matrix transformation and is a fundamental procedure in different types of text mining techniques [68]. Such transformation promotes the occurrence of rare terms and decreases the impact of more common non-stopwords. TF-IDF is separated in two parts: first, TF explains the ratio of the number of times a keyword emerges in a given document, n k (where k is keyword), to the total number of terms in the document, n: TF = n k /n and IDF is defined as follows: where N is the number of documents and N k is the number of documents that contain the keyword, K [74]. Then, TF-IDF is illustrated as follows: TF-IDF = (n k /n) * log 2 (N/N k )

Singular Value Decomposition (SVD)
Singular Value Decomposition (SVD) plays the most important role in the latent semantic analysis. SVD is a linear algebra technique, which is a factorization of a complex matrix. More information about computation of the SVD is presented in previous literature [75,76]. The major idea is to make a rectangular matrix A be broken down into the product of three matrices, an orthogonal matrix U, a diagonal matrix S and the transpose of an orthogonal matrix V, as follows: where U T U = I; V T V = I; the columns of V are orthonormal eigen-vectors of A T A; the columns of U are orthonormal eigenvectors of AA T ; and S is a diagonal m × n matrix containing the square roots of the eigenvalues from V or U in descending order. These eigenvalues indicate the variance of the linearly independent components along each dimension (factor).
Singular value decomposition gets the TF-IDF weighted term matrix and converts it into three matrices containing the term-by-factor matrix, document-by-factor matrix and singular value matrix (square roots of eigenvalues). The term-by-factor matrix describes the term loading to a specific latent factor. The document-by-factor describes the document loadings to a specific latent factor. The singular value illustrates the significance of a specific factor.
The performance of singular value decomposition in LSA in terms of simulation is almost similar to the way the human brain distills meaning in text [68]. It comes from the notion that some different words can have the same meaning or vice versa; one word in distinct contexts can have a different meaning. The words that have the same meaning will load to their common underlying concepts. This explains that one word might load multiple latent concepts in comparison to its main underlying concept. This is a key that empowers LSA in distinguishing the underlying concepts within textual data [68]. The more detailed mathematics can be found in prior studies [68,75].

Factor Determination
Singular Value Decomposition (SVD) can be calculated in different dimensions by eliminating the less important factors in the matrix. In addition, LSA can explain different levels of abstraction by decreasing the number of factors. In SVD there is a possibility to analyze data with different factors, which is usually is related to how much variance in the term vector is cumulatively explained by the factor [68,72]. However, the number of tokens in our study is too small to consider the variance and there are more specific tokens that need to be classified by the LSA. We calculated different dimensions to reach the ideal number of factors; in our study, the ideal number of the factors for the top level LSA analysis is five (Table 3). These five factors distinguish 57 tokens among 101 tokens in the word vector that specifically explain which words have been applied more.

Term Loadings and Cross-Loadings
As it is possible in exploratory factor analysis for one item to load multiple factors, in the LSA, one token might also load multiple factors. This comes from human language, in that one word may have different meanings according to context. For instance, the token "spindl" load from Factor 1 to Factor 5, respectively, as follows: 0.0295, 0.0045, −0.1682, 0.0547 and 0.1890. It shows that the token "spindl" were used in all factors but with different meanings or purposes, which made us decide to choose the high-loading term for each token in the five factors. Therefore, the token "spindl" was added to Factor 5 and the other loadings were named cross-loadings. Table 1 shows the high-loading terms after omitting the cross-loading terms (Appendix A shows all high-loading and cross-loading terms).

Factor Interpretation and Labels
The meaning of each factor in Table 1 is explained by the terms and documents loaded to it. We created a new table (Table 4) that shows how these tokens describe sub-criteria in the MCDM framework. Now it is essential to label the factors and interpret the tokens. As it was mentioned in previous studies [68,72], labeling factors in LSA can be a challenging task, as usually there are no corresponding explanations or phrases in order to match to a specific factor. Labeling the criteria was decided by the authors. In addition, the authors read all the articles separately in order to realize the relation of each token and the previous literature. In this step, the authors read each token and found it in the literature in order to find out the most fitted sub-criterion, because each token does not make sense alone, and most of the sub-criteria are a mixture of some words. In some cases, making two or more tokens up concluded a sub-criterion; but, in some cases, merging a token and some other words out of Table 1 showed us the best and the most fitting sub-criterion. The first factor category or criterion is called "Size and Precision" because the tokens are referring to the previous studies that discussed the details. For example, ultimate load capacity, which is an important sub-criterion in machine tool selection, is considered as Precision. The second sub-criterion "diversity of materials and structure" discusses the changeability of a tool in terms of choosing materials or structure. Moreover, weight machine dimensions, maximum speed, maximum tool diameter and product conformance discus the size and precision at the same time. In some previous studies [77], Size and Precision were considered as two different criteria; however, in this study we show that they are meaningfully related to each other interchangeably. The second criterion is "Cost and Serviceability". Unlike the previous studies [77][78][79] that considered cost and serviceability as two different criteria, in this we demonstrate that these two criteria are somehow cause and effect elements. Price, energy consumption, maintenance cost, waste amount, operation cost, quality of technical service, service training and repair service are sub-criteria of the Cost and Serviceability criterion. The tool's flexibility has always been an important criterion in order to select an optimum choice. In machine tool selection, the literature setup time, installation easiness, ease of learning, ease of use, integration, properties and user friendliness fitted very well with the third classification of the tokens. "Productivity" is a well-known criterion in choosing all different type of machines and technologies. In this study, the best fitted tokens in the productivity classification are as follows: depreciation quality, scrap and rework reliability, pallet changer and fixture, labor cost as well as operation shifts. The last criterion, which is also is a mixture of two distinct criteria in prior studies, is called "Technical Features and Safety". These two criteria are straightly related to each other and it is better to consider them in a unique criterion. Because they are cause and effect criteria, the better and high-quality the technical features are, the higher the safety is, and vice versa. The most fitted tokens in the fifth classification and sub-criteria in the literature are as follows: capacity of rotary table, thermal deformation, spindle speed, spindle power, adaptability, failure rate, tool changer time, fire extinguisher, number of tappers and reliability of the drive system. These sub-criteria show that the most discussed type of machine tool was Computer Numerical Control (CNC). Regarding the literature, the previous studies that discussed the sub-criteria are displayed in Table 2. These references were concluded by SVD output in RapidMiner 9.6, and the results were classified based on high-term loadings in each factor; the related methodologies in each reference are demonstrated in Table 4.

Confluence of PMADM and Text mining
As mentioned above, the application of text mining in PMADM is novel and has space to grow. The innovation in our results is that the prior studies presented each criterion separately, but the text mining approach shows that some criteria are interchangeably connected to each other based on prior literature containing Size and Precision, Cost and Serviceability, and Technical Features and Safety. The following Table 4 demonstrates the results of the Latent Semantic Analysis in finding the most discussed criteria in previous studies, which allow us to anticipate future research. In order to adjust the obtained tokens and sub-criteria, the authors needed to read all prior studies in detail to understand the relation between them.

Final Proposed Weighting Structure
In this section, each criterion and sub-criterion's weight will be calculated. As a result of the last section, sub-criteria were obtained based on a text-mining approach. With help of the LSA methodology, the most significant tokens were classified into five different categories, which are called the machine tool selection criteria. In the first step, the authors found the number of occurrences of each sub-criterion in the literature. For example, the phrase "Ultimate load capacity" were repeated in three different documents in the literature; thus, the number of occurrences were counted for each criterion and are shown in Table 5. In order to acquire each sub-criterion weight, the number of occurrences is summed in each criterion, and then every number of occurrences is divided by summation. This procedure is concluded by a decimal number that illustrates the sub-criterion weight. This procedure carries on for all criteria. For instance, the wight of the sub-criterion "Ultimate load capacity" is 0.11, which is obtained by the division of 3 by 27. To find out the criteria's weight, the last procedure was implemented in a higher level. The number of occurrences of each criterion was divided by the summation of all criteria. For example, the "Size and Precision" occurrences were 27, which was divided by the summation of all criteria, 149, with the result being 0.18. The other results are shown in Table 5.
The obtained results show us the importance of criteria in machine tool selection literature. Cost and Serviceability has the highest priority among the criteria, with 0.34 as the weight. The classification of criteria considering high priorities is identified as follows: (1) Cost and Serviceability; (2) Technical Features and Safety; (3) Size and Precision; (4) Flexibility; and (5) Productivity. Therefore, Table 5 displays the importance of each sub-criterion with its weight in order to find out the importance priority.
Generally, it is common to have a comparison between MADM methods to analyze the advantages and disadvantage of similar methods of weighting or ranking. This study presented a unique way of criteria selection and weighting, which is not based raw experts' opinions about a subject or topic.

Conclusions
As can be seen in Tables 4 and 5, the criteria selection andf weighting system was presented in a special case study, which was "Machine Tool Selection". This study showed how other researchers can apply text mining for the process of criteria selection and weighting in MADM and Prospective MADM. This new approach and model can be applied in many other cases and topics with bigger data bases. There are many proper data bases, such as Scopus, that can help to have more reliable answers and outputs for solving complicated decision-making problems. RapidMiner 9.6 is really powerful software in data and text mining, and as it was illustrated in the study, can be a convenient way to apply test mining methods for criteria selection and their weighting procedure.
In order to illustrate the importance of criteria selection and the weighting the criteria (referring to the MADM field), this study can be introduced as a new perspective in the literature review of weighting criteria, far from pairwise comparisons and policy-based decision-making models. This new approach can be applied with other newer contributions in the PMADM outline, such as "Multi-Aspect Criterion" or "Supportive-backup criteria", and still can be developed more with other tools and methods in text mining, or by adding newer items to the classic scheme of MADM in the wider area of the PMADM items and models.
Researchers in the field of MCDM can use this new framework as a new method for criteria selection and as a novel weighting system. Formerly, the MCDM field did not have any special way of criteria selection and many researchers tried to use social science methodologies to propose a proper way. From now on, this new proposed method can be applied in other decision-making problems, especially future-based ones like Prospective MADM models.