# node2vec: Scalable Feature Learning for Networks

**Authors:** Aditya Grover, Jure Leskovec  
**arXiv:** [1607.00653](https://arxiv.org/abs/1607.00653)  
**Date:** 2016-07-03  
**Summary page:** https://www.summarizepaper.com/en/paper/1607.00653v1/

## Key points

- Network analysis prediction tasks often require careful engineering of features
- Representation learning shows promise in automating prediction by learning features themselves
- Node2vec is a novel algorithmic framework for learning continuous feature representations for nodes
- Node2vec aims to map nodes to a low-dimensional space of features that preserves network neighborhoods
- Node2vec captures diverse connectivity patterns and complexity present in real-world networks
- Node2vec outperforms existing techniques in multi-label classification and link prediction tasks
- Node2vec can be applied in social network analysis, recommendation systems, and biological network analysis.

## Summary

In the field of network analysis, prediction tasks involving nodes and edges often require careful engineering of features used by learning algorithms. However, recent advancements in representation learning have shown promise in automating prediction by learning the features themselves. To address existing limitations in capturing diverse connectivity patterns observed in networks, the authors propose a novel algorithmic framework called node2vec for learning continuous feature representations for nodes. The goal of node2vec is to map nodes to a low-dimensional space of features that maximizes the likelihood of preserving network neighborhoods. This added flexibility enables the algorithm to learn richer representations that better capture complexity and diversity present in real-world networks. Node2vec is demonstrated through experiments comparing it to existing state-of-the-art techniques on multi-label classification and link prediction tasks across various real-world networks from different domains. The results show that node2vec outperforms these techniques, highlighting its ability to efficiently learn task-independent representations in complex networks. This work presents a new approach for effectively learning feature representations in networks by leveraging the power of representation learning and incorporating flexibility into neighborhood exploration with potential applications in various fields such as social network analysis, recommendation systems, and biological network analysis.