How Hulu Boosted Recommendation Diversity with Determinantal Point Processes

How Hulu Improves Recommendation Diversity

Hulu serves millions of videos, making it crucial to recommend content that not only matches user interests but also offers diverse options. Traditional recommender systems focus on accuracy metrics such as predicted ratings or click‑through probabilities, which can lead to overly similar recommendation lists.

For example, a new user who watches Harry Potter 1 might receive a list dominated by other Harry Potter movies if only accuracy is considered. A more varied list would also include items sharing genre, actors, or directors, as illustrated in Figure 1.

The diversity metric measures how dissimilar items in a recommendation list are. Higher diversity helps users discover new content and allows the system to uncover latent interests, but it often comes at the cost of reduced accuracy. Balancing these objectives remains a key challenge.

Hulu adopts Determinantal Point Processes (DPP) to enhance diversity. Originating in 1975 as the “fermion process,” DPP models repulsion between items, making it suitable for selecting diverse sets. Each item is represented by a relevance score and an embedding vector; the product of these forms a representation vector. The probability of selecting any subset is proportional to the squared volume of the parallelepiped spanned by its representation vectors, naturally favoring relevant yet dissimilar items.

Finding the subset with maximum probability is NP‑hard. While a greedy algorithm provides a good approximation, its naïve implementation requires computing a matrix determinant for each candidate addition, which is computationally expensive for online systems.

To overcome this, Hulu’s team developed an incremental update technique that efficiently computes the change in probability when adding an item to the selected set, dramatically reducing computational complexity. Experiments comparing the baseline lazy greedy algorithm, an ICML 2017 approximation, and Hulu’s fast greedy method (FaX) show that FaX is about 100× faster than the baseline while preserving the same probability scores.

The algorithm has been integrated into Hulu’s live recommendation pipeline, with multiple A/B tests confirming improved recommendation metrics. Moreover, the approach can be applied to other machine‑learning tasks that employ DPP.

References

[1] Yin Zheng, Bangsheng Tang, Wenkui Ding, and Hanning Zhou. “A Neural Autoregressive Approach to Collaborative Filtering.” ICML, 2016.

[2] Laming Chen, Guoxin Zhang, and Hanning Zhou. “Fast Greedy MAP Inference for Determinantal Point Process to Improve Recommendation Diversity.” NIPS 2018 (arXiv:1709.05135).

[3] Insu Han, Prabhanjan Kambadur, Kyoungsoo Park, and Jinwoo Shin. “Faster Greedy MAP Inference for Determinantal Point Processes.” ICML, 2017.