TLDR: Supported Abstract Argumentation for Case-Based Reasoning (sAA-CBR) is a novel binary classification model that enhances its predecessor, AA-CBR, by introducing ‘supports’ between cases. This innovation ensures that all past cases contribute meaningfully to the decision-making process, effectively eliminating ‘spikes’ (extraneous data points) and leading to more comprehensive and accurate predictions for new cases.
Researchers have introduced a new binary classification model called Supported Abstract Argumentation for Case-Based Reasoning (sAA-CBR), designed to enhance how past data informs decisions for new cases. This model builds upon its predecessor, Abstract Argumentation for Case-Based Reasoning (AA-CBR), by addressing a critical limitation: the presence of ‘spikes’.
In AA-CBR, past cases engage in a debate, arguing for their assigned labels and attacking cases with opposing labels. However, AA-CBR’s design, which constrains attacks to cases with minimal differences, can lead to ‘spikes’ – cases that exist in the dataset but do not contribute to the classification of new cases. These extraneous cases can hinder the model’s effectiveness and interpretability.
sAA-CBR overcomes this by introducing a ‘supports’ relation. This means that cases with agreeing labels can now support each other, in addition to the existing attack relationships. This seemingly simple addition ensures that all cases within the ‘casebase’ (the collection of past, labelled examples) actively participate in the debate surrounding an unlabelled new case. The model then applies argumentation semantics to determine which arguments are accepted or rejected, ultimately classifying the new case.
To establish these attack and support relationships, sAA-CBR relies on a defined partial order over the casebase, which determines a notion of ‘exceptionality’. More exceptional cases are able to attack or support less exceptional ones. A minimality constraint is also applied to ensure that relationships are only formed between the most similar cases, preventing superfluous connections.
Consider a practical example: assessing a patient’s diet based on features like fruit intake, calorie consumption, water intake, and high-fat food intake. If a new patient presents with all these features, sAA-CBR can determine if their diet is healthy or unhealthy. In a scenario where AA-CBR might have a ‘spike’ (a case that doesn’t contribute to the debate), sAA-CBR ensures that this case, by supporting another, can indirectly influence the outcome. This can lead to a different, and potentially more accurate, prediction because all available evidence is brought to bear on the decision.
The researchers have formally proven that sAA-CBR contains no spikes, provided the default case (representing a default expected outcome) is the least exceptional. This is a significant improvement, as it guarantees that every data point in the casebase contributes meaningfully to the classification process, without trading off other key model properties.
The underlying framework for sAA-CBR is a Bipolar Argumentation Framework (BAF), which extends standard argumentation frameworks by including both attack and support relations. For computational purposes, this BAF is translated into a traditional Argumentation Framework (AF) by interpreting supports as ‘supported attacks’ or ‘secondary attacks’. The ‘grounded semantics’ are then applied to determine the set of accepted arguments, leading to the final classification.
Also Read:
- Navigating AI Ethics: A Fuzzy Logic Approach to Moral Machines
- BranchNet: A New Approach to Interpretable AI for Structured Data
In conclusion, sAA-CBR represents a notable advancement in computational argumentation and case-based reasoning. By integrating the concept of supports, it effectively eliminates extraneous data points, ensuring that all available information contributes to more robust and interpretable binary classifications. Future work will involve empirical analysis of its classification performance and exploring the application of supports in other forms of AA-CBR. You can read the full research paper here.
