Rule Chains for Visualizing Evolving Fuzzy Rule-Based Systems

Sascha Henzgen, Marc Strickert, Eyke HüllermeierCORES 2013 · DOI

TL;DR

This paper introduces "rule chains," a novel visualization technique to help understand how fuzzy rule-based systems adapt and evolve over time, especially in dynamic environments with concept drift. By representing rule histories as chains of "pearls" and using different link types to show changes in antecedents and conclusions, rule chains offer a clear visual overview of model dynamics.

What Problem Does This Paper Address?

Fuzzy rule-based systems are powerful tools for modeling complex systems, particularly when data streams in dynamically changing environments. These systems learn and adapt incrementally, meaning their internal rules change over time. However, understanding how and why these models evolve can be challenging, especially for users who need to interpret the system's behavior or diagnose issues. Existing visualization methods often focus on static models or don't adequately capture the temporal evolution of rules. This paper aims to bridge this gap by providing a visual tool that specifically tracks the life cycle and changes of individual rules within an evolving fuzzy system.

Methods

The core contribution is the rule chain visualization technique. A rule chain represents the history of a single rule as a sequence of "pearls," where each pearl signifies the rule at a specific time point. These pearls are arranged horizontally, with earlier time points on the left and later ones on the right.

Key aspects of the visualization include:

  • Rule Pearls: Each pearl represents a rule at a given time.
  • Links: Pearls are connected by links that visualize the changes between consecutive time points for that rule. Three main types of links are used:
  • Antecedence Similarity Links (Ellipses): These links represent the similarity between the antecedent parts (the "if" conditions) of two consecutive rules. The size and darkness of the ellipse indicate the degree of similarity (lower similarity means a bigger, darker ellipse).
  • Angle Similarity Blocks (Rectangles): These visualize the similarity between the conclusion parts (the "then" actions) of two consecutive rules, specifically by looking at the angle between their linear functions. Color intensity indicates similarity.
  • Conclusion Shift Blocks: These highlight the difference in the y-intercept of the conclusion parts of consecutive rules. Color intensity reflects the magnitude of this difference.
  • Rule Histories: Multiple rule chains are stacked vertically to show the evolution of the entire fuzzy system. This allows for observing rule births, deaths, and overall system dynamics.

The paper also defines several similarity and distance measures as prerequisites for the visualization:

  • Antecedence Similarity: Measures the similarity between the conjunction of fuzzy sets in the antecedents of two rules.
  • Distance between Rule Centers: Uses the Euclidean distance between the centers of the fuzzy sets in the antecedents.
  • Angle Similarity: Measures the similarity between the conclusion parts (linear functions) of two rules based on the angle between their normal vectors.
  • Conclusion Intercept Difference: Quantifies the difference in the y-intercept of the conclusion parts.

Datasets

The paper uses a synthetic dataset to illustrate the effectiveness of the rule chain visualization. This dataset is designed to simulate concept drift, a scenario where the underlying data-generating process changes over time. The synthetic data is constructed by concatenating several datasets, each characterized by cluster centers, widths, sizes, and linear regression functions. The authors explicitly state that using synthetic data allows for controlled experiments and knowledge of the "ground truth," which is not possible with real-world data.

Key Insights and Ideas

  • Visualizing Temporal Dynamics is Crucial: The paper strongly advocates for visualization as a key tool for understanding evolving models, moving beyond static snapshots.
  • Rule Chains Offer Granular Insight: By focusing on individual rule histories and the changes between consecutive time steps, rule chains provide a detailed view of model adaptation.
  • Concept Drift Manifests Visually: The authors demonstrate that concept drifts, which are changes in the data-generating process, lead to observable patterns in the rule chain visualization. This includes simultaneous changes in many rules (vertical lines) and the appearance/disappearance of rules (long vertical edges).
  • Multiple Perspectives are Needed: The use of different link types (antecedence similarity, angle similarity, intercept difference) highlights that a single measure is insufficient. Combining these provides a more comprehensive understanding of rule evolution.
  • Model Complexity is Easily Grasped: The overall size of the rule system (number of rules) can be quickly assessed from the stacked rule chains, serving as an indicator of model complexity.

Key Results

The paper presents a qualitative evaluation using a synthetic dataset that exhibits three distinct concept drifts. The key result is the demonstration that the rule chain visualization effectively highlights patterns corresponding to these concept drifts. Specifically:

  • The visualization shows noticeable patterns when the fuzzy system reacts to changes in the data-generating process.
  • Simultaneous changes in many rules (due to concept drift) appear as vertical lines in the stacked rule chains.
  • The appearance or disappearance of rules due to concept drift results in long vertical edges.
  • The authors show that their visualization tool can help identify when a rule system adapts to concept drift and when it reaches a stable state.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the observed patterns in rule chains are indications of concept drift, not definitive proof. The visualization shows model evolution, not data evolution directly.
  • The current implementation is a proof of concept using a specific learning algorithm (FLEXFIS) and synthetic data.
  • Future Work:
    • Displaying rule similarities and evolving rules at the attribute level.
  • Visualizing global model state characteristics.
  • Implementing interactive control of evolving fuzzy system models.
  • Extending the tool to handle more complex time-adaptive machine learning methods beyond TSK models.
  • Investigating the display of rule similarities and evolving rules at the attribute level.

Summary updated

BibTeX
@inproceedings{henzgen2013rule,
  author = {Sascha Henzgen and Marc Strickert and Eyke Hüllermeier},
  title = {{Rule Chains for Visualizing Evolving Fuzzy Rule-Based Systems}},
  booktitle = {CORES},
  year = {2013},
  doi = {10.1007/978-3-319-00969-8\_27},
}
Figure