Laming Chen Guoxin Zhang Hanning Zhou Hulu LLC

Laming Chen, Guoxin Zhang, Hanning Zhou Hulu LLC

UNIQUE MEDIA PARTNER RELATIONSHIPS 540+ 1000+ CONTENT PARTNERS ADVERTISING PARTNERS

Recommender System in Hulu Smart Start Action Cover Stories

Recommender System in Hulu Reco Reason

Recommender System in Hulu

Recommendation Goals • Exploitation • Exploration • Relevant - users are likely to take the actions which are shown (browse series, watch content, add to watch list, . . . ) • Coverage - new content has an opportunity to surface for a user (not churning between same actions over and over). A significant portion of available content are recommended • Transparent - it’s clear why the content selected is being recommended to the user • Contextual - the content selected is timely and relevant to a person’s profile, device, and location • Adaptive - recommendation results adapt quickly to new users, new behavior, and newly available content • Serendipitous - recommending content not obviously linked to previously expressed preferences • Diverse - Don’t show the same recommendation to a user many times in a short period. Mixture of various content type, genre, or network / brand

Why Diversity? • In item-based CF models, only focusing on accuracy metrics often leads to similar items

Related Works • • • Maximum Marginal Relevance, 1998 Greedy, 2001 Topic Diversification, 2005 Absorbing Random Walks, 2007 Relaxation and Quantization, 2008 Clustering Based Diversification, 2014 • Latent Factor Portfolio, 2012 • Set-oriented Personalized Ranking, 2013 • Determinantal Point Process Eigenmixtures, 2013

Related Works • False True

Related Works • Determinantal Point Process (DPP) in Basket Completion: • Expectation-maximization, 2014 • Fixed-point algorithm, 2015 • Low-rank factorization, 2016

Our Work •

Introduction of DPP •

Geometric Interpretation of DPP • Picture from [Determinantal point processes for machine learning, Alex Kulesza, Ben

Greedy vs. DPP • Common • Direct integration into existing recommender systems • Capable to make trade-off between accuracy and diversity • Difference • Pairwise dissimilarities vs. Volume How to implement?

DPP-based Method • False True

Acceleration Technique •

Acceleration Technique •

Acceleration Technique •

Experimental Results • Efficiency (#item = 1000)

Experimental Results • Accuracy vs. Diversity • Reference algorithms: Random, MMR, Greedy • Datasets: Movie. Lens, Last. FM, Jester SUGGEST: Evaluation of item-based top-n recommendation algorithms, George Karypis. 2001

Experimental Results • Accuracy vs. Diversity

Experimental Results

Summary •

We are hiring! https: //www. hulu. com/jobs/positions {laming. chen, guoxin. zhang, eric. zhou}@hulu. com

Algorithm