Similarity measure for aggregated fuzzy numbers from interval-valued data

Justin Kane Gunn, Hadi Akbarzadeh Khorshidi, Uwe Aickelin

Abstract

This paper presents a novel methodology for computing the degree of similarity between aggregated fuzzy numbers derived from interval-valued data using the Interval Agreement Approach (IAA). Unlike traditional similarity measures that rely primarily on geometric distance, the proposed measure incorporates a multi-dimensional feature set—including area, perimeter, modified centroids, quartiles, and agreement ratios. To optimize the contribution of each feature, weights are determined through Principal Component Analysis (PCA) on synthetic datasets. The effectiveness of the measure is demonstrated through an illustrative film recommendation scenario, highlighting its potential for nuanced pattern matching in uncertain environments.

Introduction

In many real-world decision-making scenarios, human opinions are inherently imprecise. While traditional data analysis often forces these opinions into "crisp" numerical values (e.g., a rating of 8 out of 10), this approach discards the underlying uncertainty of the respondent. Interval-valued data (e.g., "between 7 and 9 out of 10") provides a more faithful representation of this ambiguity. However, comparing sets of intervals requires sophisticated aggregation techniques. The Interval Agreement Approach (IAA) has emerged as a robust method for transforming these intervals into fuzzy numbers, representing the collective agreement across a range.

Despite the utility of IAA, measuring the similarity between the resulting fuzzy shapes remains a challenge. Existing metrics often fail to capture the structural nuances—such as the "jaggedness" of a curve or the specific distribution of expert consensus—that distinguish one fuzzy set from another. This study addresses this gap by proposing a comprehensive similarity measure tailored specifically for IAA-generated fuzzy numbers.

Methodology

Feature Extraction

The proposed similarity measure is built upon six distinct attributes that characterize the geometry and statistical distribution of an IAA fuzzy number:

  1. Area: Represents the total geometric space under the membership curve, providing a baseline for the magnitude of agreement.
  2. Perimeter: Captures the complexity of the boundary. A higher perimeter relative to area often indicates a lack of consensus or a more "jagged" distribution of expert intervals.
  3. Modified Centroids (x and y): The "center of mass" of the fuzzy number. In this study, the centroid calculation is modified to ensure that the measure is equally sensitive to changes in both high-certainty and low-certainty regions.
  4. Quartiles: A five-number summary (minimum, Q1, median, Q3, and maximum) derived from the interval endpoints, capturing the statistical spread of the underlying data.
  5. Agreement Ratio: A specialized metric ranging from 0 to 1 that quantifies the internal consistency of the original data sources.

Weighting via Principal Component Analysis (PCA)

To avoid the subjectivity of manual weight assignment, the researchers employed Principal Component Analysis (PCA). By analyzing a large, randomly generated dataset of IAA fuzzy numbers, PCA was used to identify which features most significantly contribute to the variance between different data patterns. This unsupervised learning approach ensures that the similarity measure is grounded in the mathematical characteristics of the data rather than heuristic assumptions.

The Similarity Model

The final similarity score is calculated as a weighted linear combination of the differences between these six features. The resulting score is normalized between 0 (representing total dissimilarity) and 1 (representing identity).

Experimental Results and Discussion

Illustrative Film Dataset

To demonstrate the practical application of the measure, the authors developed a synthetic dataset involving 10 films reviewed by 5 critics. Each critic provided their score as an interval. The similarity measure was then applied to compare the "reception profiles" of different films.

Key findings from the analysis include:

  • Structural Similarity: Film C and Film E achieved a high similarity score of 0.9461. While their raw average scores differed, their overall "vibe"—the specific way critic intervals overlapped and clustered—was nearly identical.
  • Distinct Profiles: Film C and Film J were identified as the most dissimilar (0.3740), reflecting both their different positions on the scoring scale and their divergent shapes of agreement.
  • User Profiling: By transposing the analysis to compare critics rather than films, the model identified Critic 2 and Critic 4 as having the highest similarity (0.9706). This suggests the measure's utility in collaborative filtering and recommendation systems.

The Role of Perimeter and Centroids

A significant insight from this research is the importance of the perimeter attribute. While many models focus on area, the perimeter serves as a sensitive proxy for uncertainty; a complex, high-perimeter shape reveals nuances in expert disagreement that area alone masks. Furthermore, the modified centroid ensures that the measure remains robust even when the "mass" of the fuzzy number is concentrated in regions of low membership.

Conclusion

This study introduces a nuanced similarity measure for aggregated fuzzy numbers that moves beyond simple geometric distance. By combining structural features with PCA-derived weights, the method provides a sophisticated tool for comparing uncertain data. Future research will focus on integrating this similarity measure into established machine learning pipelines, such as K-Nearest Neighbor (KNN) classification or TOPSIS (Technique for Order of Preference by Similarity to Ideal Solution) for multi-criteria decision-making, where the accuracy of the similarity metric is paramount to the system's overall performance.

Summary updated

Abstract

This paper presents a method to compute the degree of similarity between two aggregated fuzzy numbers from intervals using the Interval Agreement Approach (IAA). The similarity measure proposed within this study contains several features and attributes, of which are novel to aggregated fuzzy numbers. The attributes completely redefined or modified within this study include area, perimeter, centroids, quartiles and the agreement ratio. The recommended weighting for each feature has been learned using Principal Component Analysis (PCA). Furthermore, an illustrative example is provided to detail the application and potential future use of the similarity measure.
BibTeX
@article{gunn2020similarity,
  author = {Justin Kane Gunn and Hadi Akbarzadeh Khorshidi and Uwe Aickelin},
  title = {{Similarity measure for aggregated fuzzy numbers from interval-valued data}},
  journal = {Soft Computing Letters},
  year = {2020},
  doi = {10.1016/j.socl.2020.100002},
}
Figure