node2vec is an algorithm for learning continuous feature representations for nodes in networks. It maximizes the likelihood of preserving network neighborhoods by using a flexible notion of node neighborhoods and a biased random walk procedure. This approach generalizes prior work based on rigid neighborhood definitions and allows for richer representations by exploring diverse neighborhoods. node2vec is evaluated on multi-label classification and link prediction tasks across various real-world networks, outperforming state-of-the-art methods by up to 26.7% in multi-label classification and 12.6% in link prediction. It is efficient, scalable, and robust to noisy or missing edges. The algorithm is semi-supervised, allowing for the use of a small fraction of labeled data. node2vec extends feature learning from nodes to pairs of nodes for edge-based prediction tasks. It uses a 2nd-order random walk approach to generate network neighborhoods and allows for flexible exploration through parameters that control the search space. The algorithm is computationally efficient and can scale to large networks with millions of nodes. node2vec provides task-independent feature representations that are effective for a wide range of network analysis tasks. It is available at http://snap.stanford.edu/node2vec.node2vec is an algorithm for learning continuous feature representations for nodes in networks. It maximizes the likelihood of preserving network neighborhoods by using a flexible notion of node neighborhoods and a biased random walk procedure. This approach generalizes prior work based on rigid neighborhood definitions and allows for richer representations by exploring diverse neighborhoods. node2vec is evaluated on multi-label classification and link prediction tasks across various real-world networks, outperforming state-of-the-art methods by up to 26.7% in multi-label classification and 12.6% in link prediction. It is efficient, scalable, and robust to noisy or missing edges. The algorithm is semi-supervised, allowing for the use of a small fraction of labeled data. node2vec extends feature learning from nodes to pairs of nodes for edge-based prediction tasks. It uses a 2nd-order random walk approach to generate network neighborhoods and allows for flexible exploration through parameters that control the search space. The algorithm is computationally efficient and can scale to large networks with millions of nodes. node2vec provides task-independent feature representations that are effective for a wide range of network analysis tasks. It is available at http://snap.stanford.edu/node2vec.