عنوان مقاله [English]
Nowadays, with quick development of telecommunication technologies, companies are able to find their uppliers, as well as their options for partnering in business, easier than before in different countries. Based on the increase in the number of potential suppliers and their information, supplier management sections in companies have a long list of suppliers to choose from, and loads of information to process. Thus, proper methods should be adopted in order to evaluate the suppliers. On useful method is creating credit levels, and ranking suppliers based on their past cooperation with a company. For implementing such a method, a tool which makes it possible to analyze a data base and find the rules and patterns, such as data mining, is needed. Data mining can be performed by several techniques and algorithms. These techniques have different applications and each of them offer the best performance in a special situation. In this study, to achieve better results, different data mining algorithms, such as; C5.0, CART, CHAID and QUEST for classification, K-MEANS, KOHONEN and TWO-STEP for clustering, GRI and APRIORI for association rule mining (ARM) and the neural network (NN) technique, for determining the effective criteria to evaluate suppliers are used. Also, the CRISP-DM framework has been used for standardizing the process of data mining. Two indexes - calculation time and model error- have been used to evaluate the algorithms performance in training and testing data sets. In this research, neural network techniques are used to determine the significant evaluation criteria. A method is proposed to weigh the selected criteria, as well as comparing them among different data mining algorithms, to find the best technique. The techniques are applied to data gathered from ISOICO, an Iran Shipbuilding and Offshore Industries Complex Company, as a case study. The proposed method has three main phases: 1. data gathering and selecting effective evaluation criteria 2. calculating combined scores, and 3. determining features of credit levels. In phase1, all available attributes of suppliers are collected in a database. Several primary processes, such as cleaning, correcting, and categorizing, are performed to prepare the information for efficient application of data mining techniques. In this phase, attributes that are not important enough are omitted or merged together for creating more effective attributes. At the end of this phase, from 40 recorded attributes of each supplier, 30 are selected or are created by combination. All 30 selected attributes in 5 categories (key criteria, reception, quality and technical, finance, and delivery of goods) are evaluated and effective criteria are identified, without human decision and only by applying neural network techniques. The criterion for effectiveness in a neural network should have at least 5% effect on selecting the winner supplier. As a result, 18 effective criteria on selecting the winner are identified. In phase 2, after determining effective criteria, the combined scores for all suppliers are calculated. In order to calculate these scores, criteria scales are assimilated and the weights of the impacts of criteria are determined. To assimilate different criteria scales, it is assumed that these scores belong to the interval of (0,100). In this approach, the criteria which are regarded in % style are not changed but, in others, the best value is regarded as 100. Other criteria get a value proportional to the number of values of the corresponding variable of ttributes. In order to determine each criterions weight, several metrics, including the companys preference, the significance of criterion in literature, the separation capability of the criterion value, dependence or independence of the criterion, and existence of similar criterion, are considered. Finally, the combined scores of each supplier are calculated by multiplying synchronized values by weight of criteria. In phase 3, firstly, suppliers information is allocated to their corresponding credit levels. For this purpose, the number of credit levels and their corresponding score intervals are needed to be determined. In the proposed method, by assuming every 25% as a credit level, 4 levels are considered. Thus, suppliers information is assigned to their credit levels by a proportion of their combined score. Having assigned the identified information, it is possible to determine features of each credit level by applying data mining techniques, creating decision tree or if-then rules. The results show that k-means techniques offer the fastest performance (0.11s) in the relevant area of study. Also, the a priori technique offers the least model errors because its result is a rule set of equal type.