# DeepWalk: Online Learning of Social Representations

**Authors:** Bryan Perozzi, Rami Al-Rfou, Steven Skiena  
**arXiv:** [1403.6652](https://arxiv.org/abs/1403.6652)  
**Date:** 2014-03-26  
**Summary page:** https://www.summarizepaper.com/en/paper/1403.6652v2/

## Key points

- DeepWalk is a groundbreaking approach for learning latent representations of vertices in a network
- It encodes social relations within a network into continuous vector spaces
- DeepWalk leverages local information obtained from truncated random walks within the network to learn these latent representations
- It treats these walks as equivalent to sentences in language modeling, capturing the underlying structure and relationships within the graph
- DeepWalk generates highly informative representations that capture the essence of social connections
- DeepWalk outperformed challenging baselines in multi-label network classification tasks on popular social networks such as BlogCatalog, Flickr, and YouTube
- It achieved F1 scores up to 10% higher than competing methods when labeled data was sparse
- DeepWalk's representations surpassed all baseline methods while utilizing only 60% less training data in certain experiments
- One notable advantage of DeepWalk is its scalability as an online learning algorithm that generates incremental results
- It is suitable for real-world applications such as network classification and anomaly detection
- Its trivial parallelizability further enhances its efficiency

## Summary

DeepWalk is a groundbreaking approach for learning latent representations of vertices in a network. The goal of DeepWalk is to encode social relations within a network into continuous vector spaces, which can then be easily utilized by statistical models. This method builds upon recent advancements in language modeling and unsupervised feature learning, extending them from sequences of words to graphs. DeepWalk leverages local information obtained from truncated random walks within the network to learn these latent representations. It treats these walks as equivalent to sentences in language modeling, capturing the underlying structure and relationships within the graph. By doing so, DeepWalk is able to generate highly informative representations that capture the essence of social connections. To evaluate the effectiveness of DeepWalk's latent representations, several multi-label network classification tasks were conducted on popular social networks such as BlogCatalog, Flickr, and YouTube. The results demonstrated that DeepWalk outperformed challenging baselines that had access to a global view of the network, particularly when dealing with missing information. In fact, DeepWalk's representations achieved F1 scores up to 10% higher than competing methods when labeled data was sparse. Furthermore, in certain experiments, DeepWalk's representations surpassed all baseline methods while utilizing only 60% less training data. One notable advantage of DeepWalk is its scalability. It operates as an online learning algorithm that generates incremental results, making it suitable for real-world applications such as network classification and anomaly detection. Additionally, its trivial parallelizability further enhances its efficiency. In summary, DeepWalk presents a novel approach for learning latent representations in networks by encoding social relations into continuous vector spaces. Its ability to generalize language modeling techniques to graphs has proven highly effective in various multi-label network classification tasks. With superior performance compared to existing methods and its scalability features making it suitable for real-world applications involving network analysis and anomaly detection; DeepWalk holds great promise for a wide range of applications.