Fuzzy Knowledge-Based Prediction Through Weighted Rule Interpolation

Fangyi Li, Ying Li, Changjing Shang, Qiang Shen

Abstract

Fuzzy rule interpolation (FRI) facilitates approximate reasoning in fuzzy rule-based systems only with sparse knowledge available, remedying the limitation of conventional compositional rule of inference working with a dense rule base. Most of the existing FRI work assumes equal significance of the conditional attributes in the rules while performing interpolation. Recently, interesting techniques have been reported for achieving weighted interpolative reasoning. However, they are either particularly tailored to perform classification problems only or employ attribute weights that are obtained using additional information (rather than just the given rules), without integrating them with the associated FRI procedure. This paper presents a weighted rule interpolation scheme for performing prediction tasks by the use of fuzzy sparse knowledge only. The weights of rule conditional attributes are learned from a given rule base to discriminate the relative significance of each individual attribute and are integrated throughout the internal mechanism of the FRI process. This scheme is demonstrated using the popular scale and move transformation-based FRI for resolving prediction problems, systematically evaluated on 12 benchmark prediction tasks. The performance is compared with the relevant state-of-the-art FRI techniques, showing the efficacy of the proposed approach.

Introduction

Traditional fuzzy systems rely on a "dense" rule base—a complete set of "if-then" rules covering every possible scenario. However, in many real-world applications, we only have "sparse" knowledge. When a new observation doesn't match any existing rules, standard inference methods fail. Fuzzy Rule Interpolation (FRI) was developed to bridge these gaps by estimating outcomes based on the nearest available rules.

A major limitation of existing FRI methods is the assumption of equal significance: they treat every input attribute (e.g., temperature, pressure, age) as equally important. While some weighted FRI methods exist, they are often restricted to classification tasks, require manual weight assignment by experts, or need additional training data that might not be available. This paper introduces a novel approach that automatically learns attribute weights directly from a sparse rule base.

Methodology

The researchers developed a three-part framework to handle prediction tasks using only sparse fuzzy knowledge:

  1. Weight Learning from Sparse Rules: The system reformulates the sparse rule base into "artificial training instances." If a rule is missing an attribute, the system fills it with "do not care" values. It then applies unsupervised evaluation methods—specifically Local Learning-based Clustering (LLC) or Laplacian Score (LS)—to determine which attributes are most influential.
  2. Weighted Distance Selection: When a new observation arrives, the system identifies the $n$ "closest" rules using a weighted distance measure. Attributes with higher learned weights have a greater influence on rule selection.
  3. Integrated Weighted T-FRI: The learned weights are integrated into the "scale" and "move" transformations of the popular Transformation-based FRI (T-FRI). This ensures that the most important attributes dominate the shape and position of the predicted outcome.

Experimental Setup

The approach was systematically evaluated on 12 benchmark prediction tasks representing a mix of real-world challenges, including 8 multivariate regression tasks (e.g., Abalone age, Concrete Compressive Strength) and 4 time series prediction tasks (e.g., Mackey-Glass chaotic series, Gas Furnace data). Data was sourced from the UCI Machine Learning and KEEL repositories.

Key Results

The proposed method was compared against conventional non-weighted T-FRI and the state-of-the-art "CC method" (a previous weighted FRI approach).

Performance comparison of the proposed method against the CC method and non-weighted baseline across various datasets.

As shown in the performance comparison, the weighted approach outperformed non-weighted versions across all 12 datasets. In a head-to-head comparison with the CC method on seven datasets, the proposed approach achieved a significantly lower normalized Root Mean Square Error (RMSE) of 0.1351 vs. 0.6472.

Detailed RMSE results for the 12 benchmark datasets comparing non-weighted and weighted (LLC, LS, AVG) FRI methods.

Statistical validation using pairwise t-tests confirmed that the improvements were significant (typically $p < 0.05$), proving that the learned weights effectively distinguish the contributions of different domain attributes. The results demonstrate that a sparse rule base contains enough structural information to reveal attribute importance without requiring massive secondary datasets.

Conclusion

By baking attribute weights into the internal mechanism of the FRI process, the authors have improved prediction accuracy while maintaining the interpretability of fuzzy systems. Future work will address the computational costs of generating artificial instances in extremely high-dimensional spaces and explore dynamic rule base enrichment to refine weights over time.

Summary updated

Abstract

Fuzzy rule interpolation (FRI) facilitates approximate reasoning in fuzzy rule-based systems only with sparse knowledge available, remedying the limitation of conventional compositional rule of inference working with a dense rule base. Most of the existing FRI work assumes equal significance of the conditional attributes in the rules while performing interpolation. Recently, interesting techniques have been reported for achieving weighted interpolative reasoning. However, they are either particularly tailored to perform classification problems only or employ attribute weights that are obtained using additional information (rather than just the given rules), without integrating them with the associated FRI procedure. This paper presents a weighted rule interpolation scheme for performing prediction tasks by the use of fuzzy sparse knowledge only. The weights of rule conditional attributes are learned from a given rule base to discriminate the relative significance of each individual attribute and are integrated throughout the internal mechanism of the FRI process. This scheme is demonstrated using the popular scale and move transformation-based FRI for resolving prediction problems, systematically evaluated on 12 benchmark prediction tasks. The performance is compared with the relevant state-of-the-art FRI techniques, showing the efficacy of the proposed approach.
BibTeX
@article{li2019fuzzy,
  author = {Fangyi Li and Ying Li and Changjing Shang and Qiang Shen},
  title = {{Fuzzy Knowledge-Based Prediction Through Weighted Rule Interpolation}},
  journal = {IEEE Transactions on Cybernetics},
  year = {2019},
  doi = {10.1109/tcyb.2018.2887340},
}
Figure