This paper proposes a graph neural network (GNN) that leverages the rich information of directed graphs (digraphs). Existing directed graph GNNs suffer from limitations in efficiency and performance stability due to complex learning mechanisms and high-quality topological dependence. To address these limitations, we propose Directed Random Walk (DiRW), a plug-and-play strategy applicable to most space-based DiGNNs and a novel directed graph learning paradigm. DiRW utilizes a direction-aware path sampler that optimizes path probabilities, lengths, and counts without weights, considering node profiles and topologies. It also integrates a node-specific learnable path aggregator to generate generalized node representations. Through extensive experiments on nine datasets, we demonstrate that DiRW outperforms most space-based methods with its plug-and-play strategy and achieves state-of-the-art performance with its novel directed graph learning paradigm.
