انتخاب تامین کننده: یک مدل ترکیبی با استفاده از تحلیل پوششی داده ها، درخت تصمیم و شبکه عصبی

As the most important responsibility of purchasing management, the problem of vendor evaluation and selection has always received a great deal of attention from practitioners and researchers. This management decision is a challenge due to the complexity and various criteria involved. This paper presents a hybrid model using data envelopment analysis (DEA), decision trees (DT) and neural networks (NNs) to assess supplier performance. The model consists of two modules: Module 1 applies DEA and classifies suppliers into efficient and inefficient clusters based on the resulting efficiency scores. Module 2 utilizes firm performance-related data to train DT, NNs model and apply the trained decision tree model to new suppliers. Our results yield a favorable classification and prediction accuracy rate.

Supply chain vendor evaluation is a very important operational decision, involving not only selection of vendors, but other decisions with respect to quantities to order from each vendor. Globalization has led to the opportunities for many to utilize sources from around the world. This, of course, introduces additional decision-making considerations. Vendor selection decisions are complicated by the fact that various criteria must be considered in the decision-making process. Dickson (1966), in one of the early works on supplier selection, identified over 20 supplier attributes which managers trade off when choosing a supplier. The criteria may have quantitative as well as qualitative dimensions. A strategic approach towards purchasing may further emphasize the need to consider multiple criteria. In the case of strategic supplier selection, Wu (in press) stressed the need not only to consider traditional criteria such as price and quality but also longer term and qualitative criteria such as “strategic fit” and “assessment of future manufacturing capabilities”. Nassimbeni (2006) surveyed 78 Italian enterprises concerning their international sourcing, finding that quality and technological content were the highest ranked criteria for vendor selection, with cost ranked only fifth. Purchasing portfolio models have been extensively studied and findings indicate that there is no simple, standardized blue print for the application of the portfolio analysis (Olsen and Ellram, 1997 and Gelderman and Van Weele, 2003). On the other hand, supplier selection requires the information about potential suppliers’ credit history, performance history and other personal information, which are often not available to the public. Therefore, data available to supplier selection often incur problems such as small dataset available to the public, missing values, inconsistent values, errors, etc. In addition, companies conducting supplier performance evaluation always have a great deal of data but lack the knowledge of the data. That is to say, these data are not fully and effectively explored and used and they cannot provide predictive functions for the future decision-making. Many quantitative models have been proposed for vendor selection in supply chains (De Boer, Labro, & Molrlacchi, 2001). Fuzzy programming was proposed by Kumar, Vrat, and Shankar (2006) to allow consideration of various levels of uncertainty. Among various quantitative methods, both DEA analysis (Wu and Olson, 2008) and data mining (DM) techniques have been presented in vendor assessment. DEA method aids the buyer in classifying the suppliers (or their initial bids) into two categories: the efficient suppliers and the inefficient suppliers. Weber has primarily discussed the application of DEA in supplier selection in several publications; see Weber and Ellram (1993), and Weber and Desai (1996). Apart from simply categorizing suppliers, Weber demonstrated how DEA can be used as a tool for negotiating with inefficient suppliers. However, classical DEA often fails to work effectively since DEA calls for restrictions on data such as the requirement of rule of thumb, no outliers and statistical noise (Li and Reeves, 1999 and Wu, 2006). DM approaches use historical data to train a sui model and make prediction of new supplier performance with the trained model. The application cases of DM approaches to supplier selection include neural networks and expert systems (Albino and Garavelli, 1998 and Khoo et al., 1998). Due to the recent development of some methods, relatively few reports of such applications can be found in the literature. Thus it is important to investigate these methods and examine their potential. DM approaches such as decision trees (DT) and neural networks (NNs) provide good tools to approximate numerous nonparametric and nonlinear problems as well as qualitative data. The main advantage of DM algorithms involves: handling of qualitative attributes; flexibility in dealing with missing information; exploitation of large data sets for model development goal by efficient computation procedures and derivation of easily understandable classification models (classification rules or trees). The main disadvantage of using DM is that most DM techniques requires the known value of the target. In the supplier selection problem, we do not know the actual class of the supplier. This can be accomplished by DEA where efficient and inefficient signals are provided to classify all suppliers. This paper presents a hybrid model to assess supplier performance using data envelopment analysis (DEA), decision trees (DT) and neural networks (NNs). The model consists of two modules: Module 1 applies DEA and classifies suppliers into efficient and inefficient clusters based on the resulting efficiency scores. Module 2 utilizes firm performance-related data to train DT, NNs model and apply the trained decision tree model to new suppliers. Our results yield a favorable classification and prediction accuracy rate. Moreover, to our knowledge, there is no work to analyze the supplier selection problem by jointly using DEA and DM approaches. It is very attractive to use DEA and DM approaches to develop an integrated model, which exhibits the advantages of both DEA and DM approaches. This route has been proved effective in our previous work in the field of efficiency analysis of financial entities (Wu, Yang, & Liang, 2006; Wu, 2006). De Boer et al. (2001) argues that not all methods are equally useful in every possible purchasing situation. Our demonstration shows a purchasing situation where purchasing information is characterized with small dataset, qualitative data, and missing values. We begin in the following section with the proposed methodology. Section 2 provides the models and methodology utilized in this paper. Section 3 gives the DEA/DM results and further discussion. Section 4 discusses handling of qualitative data and missing data. Finally, our conclusions are presented in Section 5.

This paper has developed a hybrid supplier evaluation model using data envelopment analysis (DEA), neural networks and decision trees. The model enables us to deal with the complexity and multiple criteria including intangible criteria embedded in the supplier selection problem. The model can function as both a classification model and a regression model. For either the classification model or the regression model, it generally consists of two modules. The model consists of two modules: Module 1 applies DEA and classifies suppliers into efficient and inefficient clusters based on the resulting efficiency scores. Module 2 utilizes firm performance-related data to train DT, NNs model and apply the trained decision tree model to new suppliers. Our results yield a favorable classification and prediction accuracy rate. The above results were derived from a small dataset extracted from existing literature and need further verification using a variety of actual large data. However, the results are meaningful in that this study provides the first hybrid to integrate DEA and DM techniques and demonstrate its application to supplier selection problem. In addition, the results of this study provide insight for selecting the appropriate prediction method for a small dataset problem. A promising area of future research would be in applying this approach to compare the performance of other DM methods.