recsys 2014 keynote: the value of better recommendations - for businesses, consumers, producers, and...
DESCRIPTION
A keynote at RecSys 2014: The Value of Better Recommendations - For Business, Consumer, Producer, and Society. A story, told from the Netflix perspective, of Internet TV and how recommendations systems enable the long tail, improve economics, and spread a global culture, with thoughts on objective metrics, measurement techniques, AB testing.TRANSCRIPT
Quantifying the Value ofBetter Recommendations
The Value of Better Recommendations
Stakeholders:■ Business value■ Consumer value■ Producer value■ Cultural and Societal value
Recommenders have power for great benefit,but also for harm
Use your power wisely!
Neil Hunt
1999: Started work on recommendations at NetflixGoal was to improve satisfaction, while solving the %New problem
2006-2009: Netflix PrizePublic recognition of the importance of recommenders
2007-2011: Transition to streamingComplete catalog, short supply → curated catalog, unlimited supply
2014: 300 people working on “content discovery”$150M investment
Context
Technology EliminatesConstraints on Personal ChoiceConstraints on Personal Choice Falling Away
■ Geographic “trade radius”■ Production / manufacturing■ Shelf-space■ Channel (TV, radio)
Recommenders allow access to the long-tail of choices:■ Discovery■ Evaluation
Recommenders Enable Long Tail Media
■ There are no bad shows,just shows with small audiences
■ It’s our job to find and motivateexactly the right audience
Linear TV Channel - one choice available■ Only watch what’s being broadcast■ 21 hours/week of prime time - nothing else matters
On-demand - unlimited catalog accessible instantly■ Paradox of choice:
1000s of possibilities, most not interesting
Need a custom channel for each user (50M channels):■ 20-50 personalized choices
Netflix - TV of the Future - 50M Channels
No More Commercials
… or only relevant and interesting ads(chosen by recommenders…)
Richer Storytelling
Freed of the constraints of linear TV,not all shows must be 42 minutes with a cliffhanger end
Discovery from outside a channel grid liberates the format
The same was true for novels in Dicken’s time:Pickwick Papers was published in 20 weekly magazineswith 32 pages of text (a 4-fold broadsheet)and 16 pages of advertising support
They too, were liberated by advances in technology - making books possible
Business Objectives
Why Do Businesses Invest in Recommenders?
Better Economics…■ Makes a traditional business better, or
(Netflix, Amazon, Spotify, Pandora, ...)
■ Enables new businesses not possible before(LinkedIn, GoogleNews, Instagram, Waze, Pinterest, any free service with ads, …)
Why Do Businesses Invest in Recommenders?
The Tension:■ Enhancing customer satisfaction
■ Better choices■ Shorter time to choose
■ Suggesting more profitable products■ Choices with better margins■ Advertising
■ Long-term vs. Short-term tradeoff?
Netflix Choices
All our content is licensed to a fixed fee:Each possible choice has same cost impact
We don’t sell advertising on our service. Never will.
We don’t sell our recs or data to third parties in any form.
For Netflix, it’s all about customer satisfaction
Quantifying Netflix Benefits
7B hours per quarter50M subscribers worldwide90 minutes/day average150M choices/day
Quantifying Netflix Benefits
A good choice leads to a complete viewingA poor choice leads to abandonment, and risk of cancel
10% “better” choices → +500M/month good outcomesIf 1% of those avoids a cancellation → $500M/year
Our measurement thresholds:0.1% retention improvement ($5..50M/year)0.1% more viewing per time period
Our Business Metrics
Business Value
New Trials Retention
Hours of ViewingWord of Mouth
Retention is a Blunt Measuring Instrument
Retention is a Blunt Measuring Instrument
Measuring Users-at-Threshold
Hours of Viewing
Freq
uenc
y
Medians AveragesBaseline
Measuring Users-at-Threshold
Hours of Viewing
Freq
uenc
y
Medians AveragesBaseline Higher Avg
Measuring Users-at-Threshold
Hours of Viewing
Freq
uenc
y
Medians AveragesBaselineHigher Median Higher Avg
But We’re Still Measuring the Wrong Thing...
We optimize hours of viewing…But all hours are not created equal
Implication:■ We machine-learn addictive over compelling■ Partly innoculated by also measuring retention
What signal can we find for valued hours?
What if the Retention Driver is Something Else?Avoiding Failed Sessions (user found nothing to watch)Reducing Time-to-PlayMaximizing fraction-of-content-viewedMaximizing velocity of episode consumption
Consumer Objectives
What Consumers Say...
“I don’t need suggestions, just show me the good stuff”
“Don’t hide anything - I want to evaluate it all”
Winning The Moment of Truth
■ Moment of truth■ 1-2 minutes to find something■ 20-50 chances to connect■ Or the user has moved on...
The Intrinsic/Social Spectrum
Oracle vs. Advisor
✗ ✓
Content Producer Implications
The Cliff of Conventional Media
Producers must aim for broad audience or be irrelevant
Target Audience
Con
sum
ptio
n
Too smallNo-one knowsNo-one cares
Consu
mption
matc
hes
targe
t aud
ience
Recommender Systems Level The Cliff
Economics of high-end producers is less exponential
Producers can target the audience of their choice
New producers with niche product can emerge
Target Audience
Con
sum
ptio
n
Even small audiencescan be engaged
Consu
mption
matc
hes
targe
t aud
ience
Recommender Systems Level The Cliff
Long-tail producers aren’t excluded
Much greater cultural diversity is enabled
Does Data Drive the Product?
■
■
■
■
■
■
Netflix Use of Data for Content
✓ Predict reach & hours for a project given what we know
? Give insight to choice of cast, location, etc. if requested
✗ DO NOT dictate “she has to die at the end of S2-E1”
The director’s choices matter!
Cultural and Societal Implications
Democratization of Media
The cultural implication of the media cliff is lack of access to less prominent voices, channels, products
Recommendations systems can provide the market.
Producers are stepping in to fill that niche
The Cultural Exception
■ Marketing economics drives large commercial cultureto displace local, regional, niche culture
■ France: l’exception culturelle under GATT■ Canada: cultural exemption under NAFTA
■ Recommendation systems can reduce the swamping effect of large commercial culture
■ More to gain by exporting French culture to the world than by limiting import of global culture to France
Protectionism can yield to multiculturalism
Filter Bubbles and Echo Chambers
Proposition:■ Recommendation systems reinforce existing taste,
don’t expose users to the new, unexpected or different
If this keeps users happy, it’s likely to be true
Our experience is that diversity and serendipity play a large role in delivering recommendations that win
?
Final Thoughts
We are just scratching the surface of what’s possible
We are just scratching the surface of what’s possible
We depend upon our users trusting us with their data-- they might lose that trust
We are just scratching the surface of what’s possible
We depend upon our users trusting us with their data-- they might lose that trust
We have the ability to do amazing things for cultureor distort it horribly by following a false north-star
We are just scratching the surface of what’s possible
We depend upon our users trusting us with their data-- they might lose that trust
We have the ability to do amazing things for cultureor distort it horribly by following a false north-star
Be creative, but humble, and amaze the world!