Abstract
This paper introduces 3WCAPOS (Three-Way Causal Attribute Partial Order Structure), a novel cognitive learning model that shifts formal concept analysis from simple "set coverage" to "causal coverage." By integrating Pearl's causal inference with three-way decision-making, the authors create a hierarchical structure that is not only more accurate than traditional decision trees but also significantly more interpretable for human experts. Experimental results across multiple UCI datasets demonstrate that 3WCAPOS effectively filters out pseudo-correlations and handles boundary domain impurity better than standard statistical models.
Introduction
In the field of knowledge discovery, researchers often use Partial Order Formal Structure Analysis (POFSA) to visualize how different attributes relate to one another. However, traditional POFSA relies on "set coverage"—essentially looking at how often attributes appear together. This approach suffers from two major flaws:
- Pseudo-correlation: It identifies associations that might be coincidental rather than causal.
- Lack of Purity: In complex datasets, nodes often contain a mix of positive and negative results (a "boundary domain"), making it hard to draw clear conclusions.
The authors aim to bridge the gap between statistical association and true causal reasoning, ensuring that the resulting knowledge structures reflect the "why" behind the data, not just the "what."
Methods
The 3WCAPOS model is built upon a three-step evolution of the standard attribute partial order structure, integrating causal factors and three-way decision theory into a unified framework.
1. Causal Factor (CF)
Using Judea Pearl's "do-operator" (intervention), the authors define a Causal Factor to measure the strength of the relationship between a conditional attribute and a decision. Unlike standard probability, this measures how much the decision changes when an attribute is intervened upon, effectively filtering out noise and confusion caused by confounding variables.
2. Causal Partial Order Structure (CPOS)
The traditional frequency-based hierarchy, often visualized through Attribute Partial Order Structure (APOS) layers (ranging from Macroview to Microview), is replaced with one based on the CF. Attributes with the highest causal influence are placed at the top of the structure. This ensures that the most "important" attributes, rather than just the most "frequent" ones, drive the classification hierarchy.
3. Three-Way Decision (3WD) Integration
To handle "fuzzy" data, three-way decision theory is applied. Instead of forcing a binary "yes/no" at every node, the model allows for a "boundary domain." This approach distinguishes between action-driven 3WD and trisection-driven 3WD. If a node isn't "pure" (containing only one class), the model delays judgment and adds further causal attributes to refine the classification until it reaches a pure positive or negative domain. This results in a tree-like structure where nodes represent specific attribute states and their associated class probabilities, which are systematically resolved at lower levels.
Results and Discussion
The model was evaluated using six binary classification datasets from the UCI Machine Learning Repository, including Balloons, Early Stage Diabetes Risk, and Parkinson's. These datasets are initially represented as a formal decision context before being processed by the 3WCAPOS algorithm.
Key Performance Metrics
- Accuracy Boost: 3WCAPOS improved classification accuracy by 1% to 9% compared to the classical Classification and Regression Tree (CART) algorithm across the tested datasets.
- Superior Interpretability: Qualitatively, the 3WCAPOS diagrams were found to be more "reasonable" from a human cognitive perspective. For example, in the Balloons dataset, it correctly identified the specific combination of color and size that causes inflation, whereas traditional models produced more cluttered structures.
- Mining the "Rare but Representative": In medical datasets (like the Caesarean dataset), certain conditions are rare but almost always lead to a specific outcome. Standard decision trees often prune these as outliers, but 3WCAPOS identifies them as high-causality nodes, preserving vital information for clinical diagnosis.
Conclusion
3WCAPOS successfully integrates causality into formal concept analysis, providing a more robust and interpretable framework for knowledge discovery. By prioritizing causal influence over simple frequency and utilizing three-way decisions to manage uncertainty, the model offers a more accurate reflection of real-world decision-making. While the model currently focuses on binary classification and faces computational hurdles with high-dimensional data, future work involving ensemble learning and multi-class expansion promises to further enhance its utility.
Summary updated