Abstract
Explainability is a key challenge and a major research theme in AI research for developing intelligent systems that are capable of working with humans more effectively. An obvious choice in developing explainable intelligent systems relies on employing knowledge representation formalisms which are inherently tailored towards expressing human knowledge e.g., interrogative agendas. In the scope of this work, we focus on formal concept analysis (FCA), a standard knowledge representation formalism, to express interrogative agendas, and in particular to categorize objects w.r.t. a given set of features. Several FCA-based algorithms have already been in use for standard machine learning tasks such as classification and outlier detection. These algorithms use a single concept lattice for such a task, meaning that the set of features used for the categorization is fixed. Different sets of features may have different importance in that categorization, we call a set of features an agenda. In many applications a correct or good agenda for categorization is not known beforehand. In this paper, we propose a meta-learning algorithm to construct a good interrogative agenda explaining the data. Such algorithm is meant to call existing FCA-based classification and outlier detection algorithms iteratively, to increase their accuracy and reduce their sample complexity. The proposed method assigns a measure of importance to different set of features used in the categorization, hence making the results more explainable.
Introduction
In the pursuit of "Hybrid Intelligence"—systems where humans and artificial intelligence collaborate seamlessly—explainability remains a primary hurdle. While many modern machine learning models operate as "black boxes," Formal Concept Analysis (FCA) offers a mathematically rigorous way to represent knowledge using lattices of concepts. However, standard FCA approaches often struggle with high-dimensional data and fixed feature sets. This paper introduces a meta-learning framework designed to automatically discover "interrogative agendas"—the specific subsets of features most relevant to a given task—thereby enhancing both performance and transparency.
The Problem: Limitations of Formal Concept Analysis
FCA is a powerful tool for categorizing objects based on their attributes, but it faces two significant challenges in practical machine learning applications:
- The Feature Selection Problem: Standard FCA algorithms typically operate on a fixed, pre-defined set of features. In complex, real-world datasets, it is rarely clear which features are truly predictive and which are merely noise.
- Computational Complexity: The number of possible concepts in an FCA lattice can grow exponentially with the number of features. Processing all possible combinations simultaneously is often computationally prohibitive.
To address these, the authors propose shifting the focus from a single, static lattice to a dynamic "agenda" of questions the AI should ask about the data.
Methodology: Learning the Interrogative Agenda
The core of the proposed framework is a meta-learning algorithm that treats feature selection as an optimization problem.
Interrogative Agendas as Mass Functions
The authors define an "agenda" as a cognitive filter. Using mass functions from Dempster-Shafer theory, the algorithm assigns importance (weights) to different sets of features. This allows the system to represent uncertainty and prioritize specific "lines of questioning."
The Meta-Loop and Gradient Descent
Instead of constructing one massive lattice, the algorithm considers a collection of smaller "basis" lattices, each representing a different subset of features. The final prediction for any given object is a weighted sum of the predictions from these individual lattices. The algorithm employs gradient descent to iteratively update these weights based on a loss function, effectively "learning" which feature combinations (agendas) are the most predictive for the task at hand.
Search Strategies
To manage the potential explosion of feature subsets, the paper outlines three primary search strategies:
- Small Subsets: Limiting the search space to combinations of a few features (e.g., pairs or triples).
- Expert Knowledge: Incorporating pre-defined agendas provided by human domain experts to jump-start the learning process.
- Adaptive Selection: A recursive approach that starts with single features and combines only those that demonstrate high importance, though the authors note this "greedy" approach may miss complex feature interactions like XOR relationships.
Experimental Validation
The framework was evaluated using both qualitative and quantitative methods:
- Toy Dataset (The Apple Case): The authors used a dataset of eight apple types described by attributes like color, sweetness, and price. The algorithm demonstrated its ability to resolve conflicting information—for instance, learning to prioritize "sweetness" over "color" when the two suggested different classifications.
- ELKI Toolkit: The meta-algorithm was integrated with the ELKI data mining toolkit to enhance existing outlier detection algorithms. Preliminary results indicate that the meta-learning layer successfully improves accuracy and reduces the amount of data required to reach stable conclusions (sample complexity).
Discussion and Key Insights
Agendas as Epistemic Filters
The most significant conceptual contribution is treating feature selection as an "interrogative agenda." This formalizes the idea that an agent’s goals and prior knowledge should dictate which data points it focuses on.
Explainability by Design
Because the model assigns explicit weights to specific feature sets, the reasoning behind a decision is transparent. A user can see, for example, that a classification was made because the model prioritized "Sweetness + Price" over "Color + Origin," providing a human-readable explanation for the AI's logic.
Conclusion and Future Work
The proposed meta-learning algorithm offers a path toward more transparent and efficient AI by focusing on relevant interrogative agendas. While the current framework shows promise in reducing sample complexity and improving explainability, future research will focus on transfer learning—applying learned agendas to new, similar environments—and ensuring fairness and proportionality when multiple stakeholders' agendas are involved.
Summary updated