Automatically Discovering Conceptual Neighborhoods Using Machine Learning Methods

Abstract

Qualitative spatio-temporal reasoning (QSTR) plays a key role in spatial cognition and artificial intelligence (AI) research. In the past, research and applications of QSTR have often taken place in the context of declarative forms of knowledge representation. For instance, conceptual neighborhoods (CN) and composition tables (CT) of relations are introduced explicitly and utilized for spatial/temporal reasoning. Orthogonal to this line of study, we focus on bottom-up machine learning (ML) approaches to investigate QSTR. More specifically, we are interested in questions of whether similarities between qualitative relations can be learned from data purely based on ML models, and, if so, how these models differ from the ones studied by traditional approaches. To achieve this, we propose a graph-based approach to examine the similarity of relations by analyzing trained ML models. Using various experiments on synthetic data, we demonstrate that the relationships discovered by ML models are well-aligned with CN structures introduced in the (theoretical) literature, for both spatial and temporal reasoning. Noticeably, even with significantly limited qualitative information for training, ML models are still able to automatically construct neighborhood structures. Moreover, patterns of asymmetric similarities between relations are disclosed using such a data-driven approach. To the best of our knowledge, our work is the first to automatically discover CNs without any domain knowledge. Our results can be applied to discovering CNs of any set of jointly exhaustive and pairwise disjoint (JEPD) relations.