Abstract
Supplier selection is one of the most important activities of a purchasing manager and involves the identification of important criteria to select appropriate suppliers from a pool of available ones, prioritise them and evaluate their performance. Traditionally, multi-criteria decision-making approaches (MCDM) have been utilised for this process. Supplier selection has become increasingly complex both in terms of selection criteria, as a result of expanded data collection processes, and in terms of supplier numbers, due to the effects of globalisation. This complexity has led to considering Artificial Intelligence (AI) techniques, primarily machine learning (ML),to enhance and further improve the supplier selection process. While AI and ML provide the advantage of increased overall performance and efficiency, they are not always understandable and easily adopted by humans. Novel hybrid solutions that integrate MCDM and ML have been introduced to combine their advantages and mitigate their drawbacks. The main hybrid approach proposed involves using interpretable ML algorithms combined with MCDM to solve supplier selection problems in two ways: The first hybrid model uses a decision tree (DT), an interpretable machine learning algorithm, to reduce the complexity of supplier selection in terms of the number of criteria or suppliers, followed by the Analytic Hierarchy Process(AHP), one of the most common MCDM techniques, to rank and select the best suppliers. This maximises familiarity and adaptability while using ML to boost performance. The second hybrid model begins with two MCDM methods, FUCOM to weigh criteria and TOPSIS to rank suppliers. These result in a labelled dataset that is used to train a DT model in the second stage of this hybrid. This classifier is used as the main supplier selection mechanism. The two hybrid approaches are formalised in a generalisable manner, which allows different MCDM and ML approaches to be employed in place of the ones explored in this thesis. The applicability of the proposed approaches is demonstrated through a case study of companies in the oil and gas sector in Libya. The case study involved two datasets, one containing over 20,000 submitted offers of suppliers to requests from different departments, and one recording the evaluation process and outcome for 2,300 requisitions, explaining the selection criteria and the reasons for selecting the chosen supplier.Experiments show that the first hybrid approach (DT+AHP) achieves considerable gains in precision, up to 90%, outperforming individual ML models in accuracy, precision, recall, and F1 score by over 14%, 18%, 11% and 16%, respectively. Moreover, an F1 score of 87% is achieved for both classes, showing that the approach performs well across both selected and non-selected suppliers. In the second hybrid approach (FUCOM+TOPSIS+DT), experiment results confirm superiority over approaches using only ML, with increases by more than 5% across all metrics and achieving an F1 score of 72.57%. Additionally, the use of DT, an interpretable ML approach, in combination with MCDM methods familiar to procurement stakeholders, allows for the results of both proposed approaches to be more explainable and understandable than pure ML-based approaches.
This research provides practical implications for supply chain stakeholders in supplier selection by facilitating intelligent decision-making by providing two different hybrid approaches; supplier selection solutions can be tailored to individual needs, depending on whether the ability to explain outcomes is prioritised or whether it is vital to a maximising performance by limiting inaccurate selection suggestions. Such hybrid approaches can also lead to increased adoption of intelligent technologies in supplier selection and supply chain processes in general and provide fruitful ground for further interdisciplinary research in AI and supply chains.
Date of Award | 7 Jun 2023 |
---|---|
Original language | English |
Supervisor | George Bargiannis (Main Supervisor) & Grigoris Antoniou (Co-Supervisor) |