Abstract
Traditional clustering algorithms often operate as "black boxes," assigning data points to groups without explaining the underlying reasoning. Furthermore, they frequently struggle with data that exhibits stylistic variations, assuming an identically and independently distributed (i.i.d.) structure. This paper introduces Interpretable Style Takagi-Sugeno-Kang Fuzzy Clustering (IS-TSK-FC), a novel unsupervised framework that incorporates "style matrices" into fuzzy inference rules. By learning to account for stylistic nuances, the model provides transparent decision-making and achieves superior performance in both accuracy and computational efficiency across diverse datasets, including art, fashion, and architecture.
Introduction
Most clustering tools, such as K-Means or Fuzzy C-Means, focus on distance-based metrics that can overlook subtle, non-i.i.d. stylistic patterns. For example, paintings from different eras may share subjects but possess distinct artistic styles that define their groupings. IS-TSK-FC aims to bridge the gap between interpretability and style-awareness by adapting the Takagi-Sugeno-Kang (TSK) Fuzzy Inference System for unsupervised tasks.
Methodology
The core innovation of IS-TSK-FC lies in its use of style matrices ($S_k$) within the "consequent" part of each fuzzy rule. These matrices mathematically capture the specific style of a cluster, allowing the model to differentiate between groups based on their internal structure and covariance rather than just raw feature proximity.
Key components of the method include:
- Unsupervised Rule Learning: The model learns both antecedents (the "IF" part) and consequents (the "THEN" part) entirely from unlabeled data.
- Alternating Optimization: A specialized strategy optimizes consequent vectors and style matrices iteratively, ensuring high performance.
- Regularization: Frobenius norm regularization prevents excessive complexity, allowing the model to revert to a standard TSK-FC baseline when stylistic variation is absent.
Experimental Setup and Datasets
The algorithm was evaluated using 14 standard benchmark datasets (e.g., Wine, Zoo, Heart) and 7 style-specific case studies designed to test its ability to handle complex stylistic nuances.
Case studies included:
- WikiPainting: Art styles such as Early Renaissance and Ukiyo-e.
- Hipster Wars: Fashion styles including Bohemian and Goth.
- ASD10: Architectural styles like Baroque and Postmodern.
- EEG and JAFFE: Specialized datasets for brainwave patterns and facial expressions.
Key Results
IS-TSK-FC demonstrated significant improvements over state-of-the-art interpretable models:
- Accuracy: On the SimWikiPainting dataset, IS-TSK-FC achieved an accuracy of 53.71%, compared to 36.07% for standard Fuzzy Subspace Clustering.
- Efficiency: Despite the added complexity of style modeling, the optimization strategy is highly efficient. The model processed the SimWikiPainting dataset in just 1.17 seconds, whereas traditional 0-TSK-FC required 86.33 seconds.
- Interpretability: By inspecting the "firing strength" of fuzzy rules, analysts can understand which features led to specific cluster assignments, providing a level of transparency missing in black-box models.
Conclusion and Future Work
IS-TSK-FC offers a flexible and efficient approach to clustering that respects the inherent stylistic structures of real-world data. While the model currently requires manual interpretation of what a "style" represents linguistically, future work will explore integrating deep learning (such as auto-encoders) to automate feature representation and further enhance interpretability.
Summary updated