Interesting Concept Mining With Concept Lattice Convolutional Networks

The extraction of meaningful conceptual structures is often a critical task in many scientific and engineering disciplines, as it enables a comprehensive analysis of complex data in terms of both context and content. In this paper, we introduce the Concept Lattice Convolutional Network ( $\mathcal {LCN}$ ), an efficient semi-supervised learning approach to identify actionable concepts (i.e., interesting conceptual structures) based on a scalable convolutional neural network architecture that operates on concept lattices. The $\mathcal {LCN}$ captures diverse levels of global context by employing a message-passing mechanism that incorporates local structural and conceptual information within a lattice. It also employs parameter-sharing convolutional operations as conceptual filters to efficiently discern relevant concepts amidst the irrelevant ones. Moreover, it applies consistent aggregations that maintain local consistency of labeling across concepts in the lattice. Experiments on several datasets show that $\mathcal {LCN}$ can accurately identify actionable concepts and is at least three times faster than state-of-the-art exact interestingness indices.