Applying decision tree-based model in tender evaluation: case of Technical University of Mombasa
Mandale, Samuel Kumbu
MetadataShow full item record
Unfair tender evaluation and contract award in public procurement is prevalent in Kenya. This has contributed to low quality of goods, services and projects. Successful implementation of building projects is heavily impacted by taking the right decision during tendering processes. Manning tender procedures can be complex and uncertain, involving coordination of numerous tasks and persons with different priorities and objectives. Bias and inconsistent decision are inevitable if the decision-making process is wholly dependent on intuition, subjective judgement or emotions. In making transparent decision and beneficial competition tendering, there is need for a flexible tool that could facilitate fair decision making. The purpose of this research was to present a model of an IT solution integrating the concepts of supervised machine learning techniques in the context of tender evaluation in public procurement. A dataset of 100 instances comprising of 53 positive and 47 negative examples was used to train J48 decision tree classifier to build the model. After attribute selection in a WEKA environment, 4 of the 7 attributes of the dataset were used as independent variables (inputs) namely, Experience, Capacity, Number of personnel and Professionalism. A set criteria was used to determine the values of the independent variables. The dependent variable (output) was a category class with either “PASS” or “FAIL” values. To determine the class of an entity the J48 model considers all the values of the independent variables based on set rules. This algorithm was preferred due to its relatively simple model among other benefits stated herein. The dataset from TUM was divided into test data and training data for the model. The performance appraisal of the model was based on the accuracy of the classification, the precision, recall ratio, ROC curve and the F- Measure. The model was proven to be impressively accurate with an accuracy of 91.1765 % while the precision obtained was 0.857. The recall ratio was 1 and an F-measure of 0.923.