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.
The Challenge: Breaking the Filter Bubble
Most modern recommender systems are optimized primarily for accuracy, often measured by click-through rate (CTR). While effective for short-term gains, this focus frequently leads to a "filter bubble," where users are repeatedly suggested popular or highly similar items. In graph-based recommendation, standard embedding techniques like DeepWalk and node2vec rely on random walks that naturally gravitate toward "high-degree" nodes—popular items with many connections.
This creates a significant bias: popular items are over-represented in the training data, while niche or "long-tail" items are ignored. The authors of div2vec argue that for long-term user retention and satisfaction, diversity is just as critical as accuracy.
Methodology: Inverse Degree Sampling
The core innovation of div2vec lies in how random walks are generated to ensure that unpopular items are represented as frequently as popular ones in the training data.
- Inverse Degree Sampling: Unlike traditional methods that choose neighbors uniformly,
div2vecselects the next node in a walk based on its degree. The probability of moving to a neighbor is proportional to a weighting function $f(\text{deg}(v))$. - Weighting Functions:
- div2vec: Uses $f(x) = 1/x$. This treats every node with "equity," forcing the walker to visit low-degree nodes much more frequently to compensate for their lack of connections.
- rooted div2vec: Uses $f(x) = 1/\sqrt{x}$. This acts as a middle ground, balancing the need for popularity-based accuracy with the desire for diversity.
Once these diversity-aware walks are generated, they are processed using a standard Skip-gram architecture to produce final node embeddings.
Real-World Evaluation and Datasets
The researchers evaluated div2vec using both offline benchmarks and live production environments to measure its impact on actual user behavior:
- MovieLens-25M: Used for offline evaluation of link prediction, coverage, and intra-list similarity.
- WATCHA: A major South Korean VOD streaming service. The model was tested on over 21,000 users.
- LINE Wallet Coupon: A coupon recommendation service in Japan with over 6 million users and hundreds of unique items.
Key Results and Insights
The results demonstrate that diversity does not have to come at the cost of performance. In fact, diversifying recommendations can significantly drive engagement:
- Offline Performance:
rooted div2vecachieved the highest AUC scores on MovieLens while significantly outperforming DeepWalk and node2vec on diversity metrics like Coverage and Entropy-Diversity. - WATCHA Live Test: In a two-week A/B test, the
div2veclist received 66.19% more clicks and 62.07% more plays than the node2vec baseline. Crucially, total watch time increased by up to 4.19%, a major success metric in the streaming industry. - LINE Wallet Live Test: The model improved unique user clicks (Click UU) by 6.55% and the overall CTR by 2.90%.
Conclusion
div2vec proves that "equity" in graph sampling—giving more probability to unpopular nodes—can break the "rich get richer" cycle of standard recommender systems. By ensuring the long tail of items is properly trained and recommended, platforms can improve user retention and discoverability without sacrificing the accuracy of the system.
Summary updated