div2vec: Diversity-Emphasized Node Embedding

Jisu Jeong, Jeong-Min Yun, Hongi Keam, Young-Jin Park, Zimin Park, Junki Cho

Introduction

In graph-based recommendation systems, accuracy is often prioritized at the expense of diversity. Standard algorithms like DeepWalk and node2vec learn item representations by performing random walks across user-item interaction graphs. However, these methods possess an inherent bias: they naturally gravitate toward high-degree nodes, which represent popular items. This creates a "filter bubble" where users are repeatedly recommended similar, popular content, potentially leading to long-term disengagement.

div2vec addresses this by modifying the random walk process to value the "long tail" of unpopular items without sacrificing recommendation quality.

Methodology: Diversity through Inverse Degree Sampling

The core innovation of div2vec is an unsupervised node embedding method that alters the sampling probability during random walks.

  1. Inverse Degree Sampling: Unlike DeepWalk, which samples neighbors uniformly, div2vec selects the next node in a walk with a probability inversely proportional to its degree ($1/d$). This ensures that low-degree nodes are more evenly represented in the training sequences.
  2. Rooted div2vec: To balance diversity with accuracy, the authors introduced a variant using the inverse square root of the degree ($1/\sqrt{d}$). This acts as a middle ground, acknowledging popular trends while still elevating niche content.
  3. Embedding Generation: These diversity-biased walks are processed through a standard Skip-gram model to produce 64-dimensional embedding vectors for users and items.
Figure 4: Comparison of node visit frequencies between DeepWalk, rooted div2vec, and div2vec. The div2vec variants significantly flatten the distribution, giving more exposure to low-degree (less popular) nodes.

Real-World Applications

The researchers evaluated div2vec across three environments: the MovieLens-25M benchmark, the WATCHA streaming service in South Korea, and the LINE Wallet Coupon service in Japan.

Figure 3: User interfaces for the WATCHA streaming service and LINE Wallet Coupon service, where div2vec was deployed for live A/B testing.

Key Insights

  • Diversity $\neq$ Lower Accuracy: A common misconception in recommendation research is that increasing diversity must degrade accuracy. The authors found that by better representing the long tail, the model improved personalization, leading to higher AUC scores in several tests.
  • Generalists vs. Specialists: Users with diverse tastes ("generalists") tend to have higher platform retention. By encouraging diverse consumption, div2vec supports long-term business health over short-term click maximization.

Key Results

Live A/B tests demonstrated the practical efficacy of the approach:

  • WATCHA (Video): In a two-week test, div2vec outperformed node2vec significantly. The first week saw a 66% increase in clicks and a 62% increase in plays. Crucially, total watch time increased by approximately 4%, a substantial gain in the streaming industry.
  • LINE Wallet (Coupons): Integrating div2vec features resulted in a 6.55% increase in unique user clicks and a 2.90% increase in overall CTR.
  • Offline (MovieLens): rooted div2vec consistently achieved the highest accuracy (AUC) and superior diversity metrics (Coverage, Entropy-Diversity) compared to traditional baselines.

Conclusion and Future Work

div2vec successfully bridges the gap between diversity and accuracy in node embeddings. While the current model uses fixed sampling functions ($1/d$ or $1/\sqrt{d}$), future research may explore automated methods to tune these functions based on specific domain characteristics or user-base demographics. Combining the structural role capture of node2vec with the diversity logic of div2vec remains an open challenge for future hybrid models.

Summary updated

Abstract

Recently, the interest of graph representation learning has been rapidly increasing in recommender systems. However, most existing studies have focused on improving accuracy, but in real-world systems, the recommendation diversity should be considered as well to improve user experiences. In this paper, we propose the diversity-emphasized node embedding div2vec, which is a random walk-based unsupervised learning method like DeepWalk and node2vec. When generating random walks, DeepWalk and node2vec sample nodes of higher degree more and nodes of lower degree less. On the other hand, div2vec samples nodes with the probability inversely proportional to its degree so that every node can evenly belong to the collection of random walks. This strategy improves the diversity of recommendation models. Offline experiments on the MovieLens dataset showed that our new method improves the recommendation performance in terms of both accuracy and diversity. Moreover, we evaluated the proposed model on two real-world services, WATCHA and LINE Wallet Coupon, and observed the div2vec improves the recommendation quality by diversifying the system.
BibTeX
@inproceedings{jeong2020div,
  author = {Jisu Jeong and Jeong-Min Yun and Hongi Keam and Young-Jin Park and Zimin Park and Junki Cho},
  title = {{div2vec: Diversity-Emphasized Node Embedding}},
  booktitle = {ComplexRec-ImpactRS@RecSys},
  year = {2020},
  url = {https://openreview.net/forum?id=Azm9ixiipXT},
}
Figure