Friends, Strangers, and the Value of Ego Networks for Recommendation

Amit Sharma, Mevlana Gemici, Dan Cosley

Introduction

A fundamental question in social recommendation is whether the "big picture" of a global social network is necessary for high-quality recommendations, or if a user’s immediate circle of friends—their ego network—is sufficient. This research, originally presented at AAAI 2013, explores the trade-offs between global and local social information.

Researchers found that recommendation algorithms utilizing only a user's immediate friends often perform comparably to, and sometimes even better than, those with access to the entire social graph. This phenomenon is driven by "locality of preference," where individuals close in a network share significantly more similar tastes than a random distribution would suggest.

The Global vs. Local Approach

Most social recommendation research assumes access to a global social graph. However, real-world constraints—such as API limitations, computational costs, and privacy concerns—often restrict developers to a user's direct connections. The authors compared two primary strategies using a $k$-nearest neighbor ($k$-NN) baseline:

  • Global Approach: Recommends items based on the most similar users found anywhere in the entire network.
  • Ego-centric Approach: Recommends items based exclusively on the user's direct friends.

The Power of Locality and Network Density

To understand why local information is so effective, the study introduced several metrics to quantify locality. A key finding is that ego networks are significantly denser and more similar than the global network. While a global user-item matrix might have a density of 0.01%, an ego network can be 100 to 300 times denser, making the local recommendation problem computationally easier and more focused.

Table 1: Similarity and Sparsity metrics across datasets, highlighting the increased density and similarity within ego networks.

As shown in the data, a user's friends are consistently more similar to them than a random group of strangers. This is particularly pronounced in "endogenous" interests like Twitter hashtags, which primarily spread through direct social interaction, compared to "exogenous" interests like movies, which are heavily influenced by external media.

Experimental Results

The study evaluated these approaches across three datasets: Facebook Movies, Facebook Music (Artists), and Twitter Hashtags. The results demonstrated that the value of social information varies by domain.

Figure 1: Recommender performance (NDCG) comparing Local and Global approaches across Artists, Movies, and Hashtags for varying neighborhood sizes (k).

Key Findings:

  • Twitter Hashtags: The ego-centric approach actually outperformed the global approach, achieving an NDCG of 0.24 compared to 0.22 for the full network. This suggests that for hashtags, a user's top 10 friends are better predictors of interest than the top 10 most similar strangers in the entire network.
  • Facebook Artists: The ego-centric and full network models performed equally well, both reaching an NDCG of 0.17.
  • Facebook Movies: The global model (0.11) slightly outperformed the local model (0.07), reflecting the fact that movie tastes are often driven by global marketing and trends rather than local social circles.

Conclusion

The success of ego-centric recommendation highlights the immense value of local social signals. By leveraging the locality of preference, developers can build effective recommendation systems that are more privacy-preserving and computationally efficient than those requiring global network access. Future work will continue to explore how different types of social ties and information diffusion patterns further influence these dynamics.

Summary updated

BibTeX
@inproceedings{sharma2013friends,
  author = {Amit Sharma and Mevlana Gemici and Dan Cosley},
  title = {{Friends, Strangers, and the Value of Ego Networks for Recommendation}},
  booktitle = {ICWSM},
  year = {2013},
  url = {https://openreview.net/forum?id=glwskX1Ys},
}
Figure