title: ‘Discussion 11: HULU’s Recommender Systems’
author: “Sufian”
date: “11/7/2019”
output: html_document
Problem Statement:
How Hulu builds its industry leading recommendation engine
Rpub links: http://rpubs.com/ssufian/547898
Scenario Design analysis
- Who are HULU’s target users?
The average age of a Hulu subscriber is 31.Having a large portion of young people as its subscribers is
an advantage for Hulu as they can retain them for a long time into the future.
For a perspective, this is almost 25 years younger than the average broadcast TV viewer who is 55.
Likewise, there’s an average income of $92,000 in households where Hulu is watched. In fact, 85% of U.S.
millennials subscribe to OTT services.
- What are HULU’s key goals?
With the fierce competition in the online streaming industry, companies are spending money to acquire
subscribers. The strategy of online TV companies is that they’ll attain profit when they have a stable
and loyal user base. Since Hulu’s is a “small” player in an ever increasingly competitive streaming
market, its expected to lose even more money before attaining any stability. Therefore Hulu’s key goals
are two fold: attaining more subsriber base ASAP in order to attain financial stability.
- How HULU accomplish those goals?
From a user experience perspective: Hulu offers an extensive subscription video-on-demand library of
movies, live TV package as well as TV shows. This awesome blend of options is what differentiates Hulu
among other digital bundles like DirecTV and Sling TV, which offer live programming only [6].
For instance, having Emmy award wining shows like Handmaid’s Tale was one of their most sucessful effort
in garnering support and attract eyeballs to their site
From a technical perspective: Hulu uses an enhanced kind of collaborative filtering with features such
as Personalized Masthead, Watchlist and Top Picks [4], see figure below:
Hulu has also been using Matrix Factorization, such as SVD++[1], PMF[2], Local Low-Rank Matrix
Approximation (LLORMA) [3], etc. with the impressive successes.
Reverse Engineering
The Hulu online streaming already has a very solid and structured platform which makes it very difficult
for a novice like myself to reverse engineer; for sure they have smarter engineers than me.
Having said that, there are some noticeable aspects that I noticed their recommending system
is heavy on. Like most recommenders, it will show you the same kind of content that you’ve been
watching over and over again, even in its subsections.
That is fine and the norm in this industry but what if a viewer likes variety? And
the viewer like to have more variety/diverse recommendations of contents?
On a high level, I would reverse engineer this specific portion by inserting an added feature in
conjunction to its content and model filtering algorithms by keeping track of the number and types of
deviations. When the number of deviations becomes high enough or over a certain threshold level, than
it becomes a habit or an intent on the part of its viewers; movies they are starting to appreciate more.
Like an inflection point into a new trajectory within their trendline. These deviations will be
added to its similairty list and become part of the recommendations in the subsections going forward.
Potential Recommendation from Hulu’s engineering team
However solid or great their current system is, there is still space for improvement for broader
application in the industry, in terms of both efficiency and accuracy. And in 2016, Hulu’s researchers
invented a novel neural network based collaborative filtering approach, called neural
autoregressivedistribution estimator for collaborative filtering (CF-NADE) [5], which is
state-of-the-art on several public benchmarks. This work was accepted and presented in the 2016
International Conference for Machine Learning (ICML 2016).
CF-NADE has shown to be a successful application of deep learning in recommendation systems, it is the
state-of-the-art with explicit feedback, where users give each item an explicit rating. While in
practice, explicit feedback is rare, but implicit feedback like watch/browse/search/purchase behaviors
are abundant. Adapting CF-NADE to implicit feedback will be the focus in Hulu’s near future.
Reference
[1] Koren, Yehuda. “Factorization meets the neighborhood: a multifaceted collaborative filtering model.”
Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM,
[2] Mnih, Andriy and Salakhutdinov, Ruslan. Probabilistic matrix factorization. In Advances in neural
information processing systems, pp. 1257–1264, 2007.
[3] Lee, Joonseok, Kim, Seungyeon, Lebanon, Guy, and Singer, Yoram. Local low-rank matrix approximation.
In Proceedings of The 30th International Conference on Machine Learning, pp. 82–90, 2013.
[4] Salakhutdinov, Ruslan, Mnih, Andriy, and Hinton, Geoffrey. Restricted boltzmann machines for
collaborative filtering. In Proceedings of the 24th international conference on Machine learning, pp.
791–798. ACM, 2007.
[5] Yin Zheng, Bangsheng Tang, Wenkui Ding, Hanning Zhou. A neural autoregressive approach to
collaborative filtering, In Proceedings of the 33th international conference on Machine Learning, NYC,
[6] https://muchneeded.com/hulu-statistics/
getwd()## [1] "C:/Desk top Stuff/CUNY Fall 2019 Folder/Data Acquisition/Project Final/Discussion 11"