Introduction
As datasets expand into the "Big Data" era, traditional fuzzy inference systems (FIS) encounter significant bottlenecks. Models like the Mamdani-type system often suffer from high computational complexity, making them impractical for massive datasets. Furthermore, data is frequently fragmented and distributed across various locations. Independent clustering of these subsets often leads to "mismatched" results—where a cluster in one subset represents a different feature than a cluster in another—preventing the creation of a unified, accurate model.
This paper introduces a novel framework that integrates Takagi-Sugeno-Kang (TSK) fuzzy systems with Preprocessed Collaborative Fuzzy Clustering (PCFC). By partitioning large datasets and employing a unique "mapping" mechanism to align features, the proposed model achieves superior performance while requiring significantly less training data.
Methodology
The proposed architecture follows a multi-stage process designed for efficiency and alignment:
1. Data Partitioning and Initial Clustering
Large datasets are divided into equal subsets. Standard Fuzzy C-Means (FCM) clustering is initially applied to each subset to generate cluster centers (prototypes) and partition matrices, which define the degree of membership for each data point.
2. Collaborative Phase and Mapping
The core innovation lies in the "Phase II" mapping technique. Before subsets collaborate, the system aligns clusters by rearranging the rows of the partition matrices. This ensures that similar properties across different subsets are matched correctly, acting as a "Rosetta Stone" for distributed data. Following alignment, the subsets share structural findings (information granules) to refine a global model.
3. TSK Fuzzy Inference System
The refined clusters are integrated into a TSK-type model. Unlike Mamdani systems, TSK uses functional outputs (typically linear equations) for the consequent ("THEN") part of fuzzy rules. This approach is computationally efficient and highly accurate for modeling dynamic systems.
Experimental Evaluation
The model was benchmarked using two nonlinear dynamics system identification problems:
- Problem I: A complex plant recognition task with three inputs and one output (1,000 data patterns).
- Problem II: A nonlinear system defined by a sinusoidal function (400 data patterns split into two subsets).
The results were compared against the standard Matlab-based Genfis2 method.
Key Insights
- Alignment Necessity: Without the mapping phase, collaborative clustering often fails due to mismatched indices from random initialization.
- Privacy Preservation: Collaboration occurs at the level of information granules (prototypes) rather than raw data, offering a potential pathway for secure multi-organizational modeling.
- Efficiency through Sparsity: Remarkably, the model achieved higher accuracy using only 50% of the training data compared to standard methods using 100%. This suggests that collaborative knowledge sharing is more effective than brute-force training on monolithic blocks.
Results
In the second nonlinear problem, the proposed method significantly outperformed the benchmark:
- Proposed Method: Training RMSE of 0.0068; Testing RMSE of 0.0070.
- Genfis2 (Benchmark): Training RMSE of 0.0134; Testing RMSE of 0.0135.
Surface plots generated by the proposed FIS demonstrated smoother and more accurate approximations of target dynamics compared to standard FCM approaches.
Conclusion and Future Work
While the results on mathematical benchmarks are promising, future research will focus on testing the system with "noisy" real-world big data, such as social media streams. Additionally, the authors aim to investigate the scalability of the mapping mechanism across dozens of distributed sites and further refine visualization tools to help human operators interpret the resulting knowledge base.
Summary updated