A Clustering Method with Historical Data to Support Large-Scale Consensus-Reaching Process in Group Decision-Making

Kai Xiong, Yucheng Dong, Sihai ZhaoInt. J. Comput. Intell. Syst 2022 · DOI

TL;DR

This paper introduces a novel framework (CMHP-LCRP) for large-scale group decision-making (LSGDM) that leverages historical preference data to improve consensus-reaching. By clustering decision-makers based on their past preferences, the method aims to achieve higher consensus more efficiently than traditional approaches.

What Problem Does This Paper Address?

Large-scale group decision-making (LSGDM) involves many stakeholders, making it challenging to reach a consensus. Traditional consensus-reaching processes (CRPs) often struggle with the sheer number of participants and the dynamic nature of their preferences. While clustering has been used to group decision-makers (DMs), existing methods typically consider only the most recent preferences, failing to capture the evolution of DMs' opinions over time. This can lead to suboptimal clustering and slower consensus. This paper addresses this by proposing a method that utilizes the historical preference data of DMs to form more robust and dynamic clusters, thereby improving the efficiency and effectiveness of the consensus-reaching process.

Methods

The proposed CMHP-LCRP framework consists of three main parts:

  • Clustering Process: This is the core innovation. Instead of using only the latest preference data, an extended k-means clustering algorithm is employed to group DMs based on their historical preference relations across all decision rounds. This historical data is transformed into vectors to facilitate the clustering.
  • Consensus Process: This involves two steps:
  • Consensus Measure: After clustering, the consensus level among DMs is calculated. This includes measuring cluster weights, preference relations within clusters, and similarity between clusters.
  • Feedback Adjustment: If the consensus level doesn't meet a predefined threshold, a feedback mechanism is activated. This mechanism identifies specific alternatives, pairs of alternatives, clusters, and DMs that need to adjust their preferences. Suggestions are then provided to guide these adjustments, aiming to bring their preferences closer to those of the best-performing cluster.
  • Selection Process: Once a satisfactory consensus level is reached, this process aggregates the collective preferences to derive a final ranking of alternatives.

Datasets

The paper uses simulated datasets for evaluation. Initial preference relations are randomly generated and then transformed into historical preference formats for the experiments. No specific real-world datasets are mentioned or used.

Key Insights and Ideas

  • Historical Data is Key: The central idea is that past preferences provide a richer, more comprehensive picture of a DM's behavior than just their latest input. This richer data leads to more stable and representative clusters.
  • Dynamic Clustering: Unlike static clustering in many existing models, the CMHP-LCRP framework implicitly allows for cluster evolution as preferences change, though the paper focuses on using historical data for initial clustering in each round.
  • Targeted Feedback: The feedback adjustment mechanism is sophisticated, identifying not just which DMs need to adjust, but also which preferences (alternatives/pairs) are problematic and which clusters are most influential, allowing for more precise guidance.
  • Efficiency Gains: The use of historical data and targeted feedback is shown to significantly speed up the consensus-reaching process and increase the success rate.

Key Results

The paper demonstrates the effectiveness of the CMHP-LCRP framework through simulation experiments. Key findings include:

  • Faster Consensus: The CMHP-LCRP framework consistently reaches consensus in fewer rounds (lower AZ values) compared to traditional LCRP models. For instance, it typically takes an average of 2-3 rounds.
  • Higher Success Rate: The framework achieves a higher success ratio (higher AS values, often close to 1) in reaching consensus within a set number of rounds.
  • Parameter Sensitivity:
    • Increasing the number of DMs (m) and alternatives (n) generally decreases the number of rounds (AZ) and increases the success rate (AS).
  • Increasing the number of clusters (Q) tends to increase the number of rounds (AZ) and decrease the success rate (AS), suggesting that a moderate number of clusters is optimal.
  • A smaller feedback adjustment parameter (β) leads to more DMs adjusting preferences, resulting in faster consensus (lower AZ) and higher success (higher AS).

Limitations and Future Work

  • Preference Representation: The current model primarily uses additive preference relations. Future work could explore other representations like multiplicative or linguistic preference relations.
  • Non-Cooperative Behaviors: The paper acknowledges that DMs might exhibit non-cooperative behaviors (e.g., providing dishonest opinions). Future research could incorporate mechanisms to detect and manage such behaviors.
  • Social Network Analysis: The influence of social networks among DMs is recognized as important. Extending the model to incorporate social network analysis could further enhance decision support.
  • Real-World Data: While simulations are used, applying and validating the framework with real-world LSGDM datasets would be a valuable next step.

Summary updated

BibTeX
@article{xiong2022clustering,
  author = {Kai Xiong and Yucheng Dong and Sihai Zhao},
  title = {{A Clustering Method with Historical Data to Support Large-Scale Consensus-Reaching Process in Group Decision-Making}},
  journal = {Int. J. Comput. Intell. Syst.},
  year = {2022},
  doi = {10.1007/s44196-022-00072-x},
}
Figure