An Argumentative Explanation of Machine Learning Outcomes1
Stefano Bistarelli, Alessio Mancinelli, Francesco Santini, Carlo Taticchi · Frontiers in artificial intelligence and applications · 2022
The black box model used in Machine Learning is considered one of the major problems in the application of Artificial Intelligence techniques [1] as it makes machine decisions non-transparent and often incomprehensible even to experts or developers themselves.In this paper, we provide an argumentative interpretation of both the training process and the results predicted.The goal is to build a Bipolar Argumentation Framework (BAF) [2] showing the dialectical reasoning behind the assignment of a certain class to a given record.Since we make assumptions neither on the dataset nor on the algorithm used, the presented procedure can be applied to existing models without the need for further adjustments.To illustrate our proposal, we use the Titanic dataset from www.kaggle.com,which contains records relating to people involved in the Titanic disaster.We consider three categorical features, namely Survived (the class to predict, with value 1 if the person survived or 0, otherwise), Pclass (ticket class among 1, 2 and 3) and sex (0 for woman and 1 for man), and two numerical features: Age (passenger age, ranging from 0.17 to 76) and Fare (passenger fare with values from 0 to 512).In the following, we describe the step our procedure goes through in order to find an explanation for the class Survived=1.Dataset Clustering.In the first step, starting from the input dataset, we create a new clustered dataset in which numerical features are split into categories that group ranges of values to obtain a more appropriate and concise explanation. BAF Generation.Then we build a BAF based on the correlation matrix computed among the features.By construction, the obtained BAF only has symmetric relations. Breaking Complete Symmetry.Given the correlation matrix, we apply a procedure that removes symmetric edges from the BAF to establish a causal relationship between features.In particular, we use the conditional probability [3] computed for arguments which attack/support each other.We choose the minimum values possible that keep the graph connected. Computing Extensions.To identify the set of arguments which are more likely to be accepted, we compute the semi-stable extensions [4] of the previously obtained 1 This work has been partially supported by: GNCS-INdAM, CUP E55F22000270001;