Research on Preference Disaggregation Based Multiple Criteria Decision Analysis Methods with Complex Decision Elements


Student thesis: Doctoral Thesis

View graph of relations


Related Research Unit(s)


Awarding Institution
  • Qingpeng ZHANG (Supervisor)
  • Xiuwu Liao (External person) (External Supervisor)
Award date12 Oct 2021


Multiple criteria decision analysis (MCDA) is a family of approaches supporting the decision maker (DM) to choose, rank or sort a limited number of alternatives that are evaluated by multiple criteria. These approaches are widely used in politics, the economy, and society. Traditional MCDA approaches require the DM to provide direct preference information about model parameters, which increases the cognitive burden of the DM. The MCDA approaches based on the preference disaggregation paradigm only require the DM to provide holistic judgment about some reference alternatives, and use regression techniques to infer the preference model's parameters, which effectively reduces the cognitive burden and encourages the DM to participate in the decision-making process. Therefore, such approaches are widely developed and applied to many scenarios. However, these approaches have made strong assumptions for decision elements: (i) the criteria should be monotonic; (ii) the criteria are preferential independence; and (iii) the criteria structure and criteria values should be pre-defined. With the changes in decision scenarios, the decision elements are becoming more and more complex. These assumptions cannot appropriately describe the characteristics of the decision scenarios and elements, making it challenging to handle real decision problems. For example, some financial indicators may not monotonically affect the business status of a firm and co-affect the predictions. When analyzing the consumers' preferences on an e-commerce platform, it is difficult to directly pre-define the involved criteria. In summary, research on the MCDA approaches for dealing with complex decision elements is still in its infancy. For this reason, this dissertation, based on summarizing the existing literature on preference disaggregation methods, carries out researches from three relevant perspectives. The main work and contributions are as follows:

(i) A progressive multiple criteria approach in the presence of non-monotonic criteria is proposed for multiple criteria sorting (MCS) problems. The preference model is based on piecewise linear functions and adapted to describe the DM's non-monotonic preference. Two interactive algorithms are proposed to revise the inconsistent preference information by progressively adjusting the complexity of the value functions or changing some assignment examples, and allow the DM to add extra preference information to guide the DM to assign non-reference alternatives. Compared with the existing methods for non-monotonic criteria, the proposed approach is the first research that focuses on MCS problems and enriches the methods for modeling non-monotonic preference. Moreover, the proposed approach balances the trade-off between model complexity and capacity for restoring preference information by constraining the total variations in the slopes of the marginal value functions, which helps to obtain more robust assignments. At last, the proposed approach progressively aids the DM to modify inconsistent preference information, adjust preference model, and determine the value functions through an interactive decision-making process. It helps the DM understand the preference better and make decisions when facing complex decision scenarios and elements.

(ii) To handle non-monotonic and interacting criteria, this research integrates the value function-based preference disaggregation approaches of MCDA to the machine learning models in a novel way and proposes two hybrid models based on the polynomial functions and the piecewise linear functions respectively. The preference model's parameters determining the value functions, which quantify the detailed relationships between criteria and predictions, are learned through neural networks. The proposed hybrid models relax the assumptions about preference independence and monotonicity of criteria and account for higher-order interacting criteria. Therefore, the proposed models can adapt to more complex decision scenarios and improves the applicability of the preference disaggregation methods. As a consequence of combining preference disaggregation methods and deep learning, the ability to analyze complex data and the predictive performances are improved, while the interpretability of ``black-box'' models is also enhanced. Moreover, the proposed models are extensible. They can adapt to different decision problems by replacing the neural networks or adjusting the preference aggregation model.

(iii) A data-driven multiple criteria method is proposed for the decision-making problems where the criteria information cannot be pre-defined. The proposed method first analyzes online information through using text mining techniques and discussing with the DM to define relevant criteria, determine the relative importance of the criteria, and quantify some qualitative criteria. Two indices, consumer preference index (CPI) and rank acceptability index (RAI), are then defined to help the DM revise inconsistent preference information, recommend alternatives, and analyze preference. The proposed method is the first data-driven multiple criteria method that does not require pre-defined criteria. It helps the DM accomplish tasks such as product recommendation and consumer preference analysis in more complex and realistic decision scenarios. Unlike the existent research, the proposed method uses text mining techniques to extract knowledge from unstructured data. Hence, the criteria structure and values are determined by actual data and the DM, which reduces the subjectivity and arbitrariness of the determination of criteria and helps to understand decision problems and preferences in depth. In addition, the proposed method can obtain additional information given updated online data, and it can analyze consumer preferences in different periods and scenarios. Therefore, the method can handle dynamic data.