Monotone Submodular Meta-learning under the Matroid Constraint

Meta-learning, or learning to learn, has attracted a lot of attention in the field of artificial intelligence and machine learning in the past few years. The key idea of meta-learning is to utilize prior experience and data to improve performance on new tasks. There are already a large number of meta-learning formulations applied to continuous domains. Recently, Adibi et al. [1] devised a discrete meta-learning framework called submodular meta-learning to reduce the computational cost for new tasks and each task is regarded as a monotone submodular function maximization problem with the cardinality constraint. Motivated by their framework, we study the submodular meta-learning problem with the matroid constraint in this paper. Our method utilizes prior tasks to train an initial solution set that can be rapidly adapted to future tasks. We design a deterministic two-state algorithm using the greedy idea and achieve a 1/2-approximation ratio.