Concept-wise granular computing for explainable artificial intelligence

Artificial neural networks offer great classification performances, but their internal model works as a black box. This can prevent their outcomes to be employed in real-world decision-making processes, e.g., in smart manufacturing. To address this issue, the neural network should provide human-comprehensible explanations for their outcomes. This can be achieved by exploiting domain concepts and measuring their importance for the classification. To this aim, we implement an information granulation process via a neural network specifically trained to represent data instances featuring the same (different) concept’s item close to (far away from) each other. By combining the representations for each concept, we obtain the so-called conceptual space embedding. The classification is obtained by processing it via a neural network classifier. The conceptual space embedding (i) organizes the data instances according to their concepts-wise proximity, resulting in a very informative data representation; this translates into greater classification accuracy if compared to a concept-wise approach from the state-of-the-art; and (ii) encodes each concept in one of its parts; this enables the measurement of the importance of one concept by manipulating the corresponding part of the conceptual space embedding. The proposed approach has been tested with real-world data from smart manufacturing.