| This dissertation integrates two important fields of information technology, data mining and data envelopment analysis (DEA), to provide a new tool for measuring the performance of decision making units (DMU). The DEA is a powerful performance measurement methodology for assessing the relative efficiency of DMUs. This methodology determines the efficient and inefficient DMUs in order to gain valuable information for making further improvements such as identifying the savings in expenditures and the best suitable way to distribute services which will eventually improve the productivity of the entire system. There are two typical assumptions in the DEA: (1) the DEA assumes that all DMUs are homogenous and identical in their operations, and (2) the DEA is deterministic and that leads to inaccurate efficiency assessment in the presence of outliers or unusual observations.; Many investigations have dealt with the DEA models, but few have focused on heterogonous DMUs, outlier detection, and scalability over large datasets. In this dissertation, a comprehensive model is presented. We introduce a new constraint-based clustering method for early detection of outliers to evaluate the performance scores of non-homogenous DMUs. In this method, DMUs dissimilar to the DMU under evaluation are labeled as outliers and are excluded from the analysis. This work removes the extra effort needed to predefine the dissimilarity parameters or the number of DMUs to be excluded.; Experimental results of our approach show big improvements in assessing the transportation system funding for school districts in the state of North Dakota. An extensive analysis is provided to show the characteristics of our method and how it compares with different models in terms of the quality of results. The performance of these school districts is measured several times using different economical models to get the most suitable view of the situation.; The dissertation starts with the investigation of the parametric and non-parametric performance measurements along with advantages and shortcomings of these metrics. Then, a detailed analysis of outlier detection algorithms in data mining is provided. Finally, a method called the clustering-based DEA is developed. |
