Why Netflix walked away from personalization

netflixhomepage

 

In 2006 Netflix offered a $1 million prize for anyone who could improve its movie preference recommendations by 10%. Netflix, at the time, made most of its money sending DVDs in the mail to users’ homes (Internet streaming had yet to take off), and personalization offered two major advantages as customers built their “movie queue” on the Netflix website. First, if the recommendations seemed to make sense, Netflix consumers would be happy as they searched for films online. And second, once the DVDs came in the mail, users might actually enjoy the movie — since a truly personalized prediction would be more likely to meet your taste than your guess based on a movie’s cover image and brief description. Happy ordering and happy watching built Netflix customer loyalty.

To spur improvement, Netflix did more than offer big bucks in the competition. It made public a dataset of 100 million-plus ratings on 17,000 movies, which included the customer rankings from 1 to 5 stars and the sequence in which customers watched films, and allowed competitors to play with the data. The cleverest part was a subset of the data was hidden blind, and Netflix would run the proposed new algorithms against that to see if the prediction models matched how customers really behaved in film rankings.

Mathematicians went wild. The competition was lauded by business pundits as an example of crowdsourcing genius. Because this was damned hard math, the project took years. And then in 2009, a team of mathematicians called “BellKor’s Pragmatic Chaos” actually cracked the code, achieved a 10% lift, and Netflix gave them the $1 million.

And then … Netflix never implemented the winning algorithm. Because personalization at that point no longer mattered.

What happened? Netflix at the time said the technical work of implementing the new personalization would be too costly for the anticipated return. These seems like a rather lame excuse, since bundling a new math model into a computer system surely doesn’t cost more than a bit of coding. Other observers noticed that, by 2010, Netflix’s business model had changed, moving away from DVDs-by-mail to instant streaming. When you can order any movie instantly online, personalization isn’t as valuable — since if the movie is a dog, you simply click over to another movie. Today, in 2014, Netflix’s online interface has a series of rows of film titles, and most of them aren’t personalized recommendations at all.

You learned my needs. But I don’t really care.

The deeper issue is that personalization is not as exciting as many once believed. In the 1990s, Don Peppers built a consulting business on the concept of “1to1 marketing,” where new computer systems would learn individual preferences and businesses would respond with customized offers. Don’s concept was that personalization would create an unbreakable competitive advantage — because once a consumer trained a company to anticipate her needs, she would be reluctant to go through the same process with a competitor. Don was observant enough to note that such customization wouldn’t be a fit for every business model — but companies that had customers with a wide range of needs (such as Netflix movie watchers) or a wide range in value (say, financial advisors courting investors) would benefit by deploying 1to1 personalization.

Despite the noble dream of giving customers more utility and companies more brand loyalty, personalization never took off. Amazon was really the best case study … but it struggles still to offer truly relevant personal recommendations on its website (the core challenges being it cannot easily recognize multiple users on the same Amazon account, or differentiate between your modality as you shop for your spouse one day and yourself the next). Twitter has a personalization engine behind its “Discovery” tab to push news or links to you based on your observed Twitter profile. That site section has so little utility, most Twitter users don’t use it. And Facebook, which arguably has the greatest trove of data on human personal interests, is really at the mercy of the advertisers who wish to target you; this is why you, guys, get ads for men’s underwear whether you really want them or not.

Why is personalization so difficult? Why is it so hard to anticipate what people want, and use that for business advantage? The challenge is personalization is at odds with a core driver of consumer purchase behavior — novelty. Consumers are constantly hungry for something new, something improved, something that will stimulate their endorphins in a manner unseen before. The iPhone 5S had marginal improvements over the prior models, but people lined up in droves for the new OSX, excited by — wait for it — a thinner Helvetica font. Most cable on-demand movie rentals are “new releases,” when logically the utility you derive from a film should have little to do with whether it was released in 2013 or 2003. Retail stores make a business of rotating in new fashions that don’t keep your body any warmer, but spring your desire to shop.

Psychologists have termed our love for newness “the novelty effect,” and it has both positive and negative implications. Humans have the highest stress response when first faced with a threat, likely a survival mechanism that spurs our fight-or-flight reflexes when a mountain lion appears over the hill (and which also explains our grumpiness on Monday mornings at work), so new things can make us angry or upset. But people also have the highest interest when a new person enters their life, a new service is launched, or a new technology is offered. Teachers have noticed, for instance, that when educational information is presented in a new medium — say, tablets — students’ test scores initially rise; the information could be the same as that presented in a history book, but the novelty of the new approach lifts interest and recall.

There is also an evolutionary bias toward novelty in human relationships, both emotional and sexual. A 2012 study by The Journal of Marital and Family Therapy found that at least one spouse in 41% of marriages admitted to marital infidelity. The reason, psychologists believe, is that physical and emotional excitement often diminish in long-term relationships. This could be tied to an ancient instinct for humans to be sexually promiscuous to spread their genes as randomly as possible, ensuring the survival of our species. Even with love, people will trade proven history for risky novelty.

And that is the trouble with personalization. Finding something new is likely at odds with our old interests — because newness by definition is a break from the past. It’s an interesting lesson for marketers now playing with big data. Yes, you can learn and model the past needs of your customers. But just as Netflix didn’t implement a 10% improvement in personalized recommendations because its analysts didn’t predict much value from it, you may find that launching a new product with different sheet metal or miracle fiber gets customers more excited tomorrow than a personalized version of what they wanted yesterday.

No one ever asked for a phone with a camera on it. But today, we can’t live without it.

 

 

10 thoughts on “Why Netflix walked away from personalization

  1. While I agree with the gist of the article I disagree with some of the details.

    I do care about the “newness” of a movie, but not because older movies aren’t “novel”. There’s simply a 99% chance I’ve either already seen the movie or that I decided (years ago) that I’m not interested in watching it.

    Netflix isn’t just lazy when it comes personalisation, it’s annoyingly anti-personalisation. It insists on showing movies in the overview which I’m not interested in watching without giving me any way to hide them.

  2. An excellent article. However, I believe that there is another equally powerful reason why personalization has failed…and it has nothing to do with novelty. Privacy is king. Even before Snowden’s revelations, some of us are so protective of ourselves (in this new age of social media) that we have never joined Facebook nor made a single recommendation on Amazon. Is it possible to be so overly concerned with privacy that it begins to border on paranoia? Certainly. But on the other hand, you have to wonder how would the world be transformed if 100% correct personalization apps were made available to all. Just imagine if every telemarketer knew all of your porn habits, or if every insurance agent could pull up your complete purchase history of alcohol and tobacco products. The sheer volume of advertising that we endure today is invasive enough…why enable the offending corporations? If I found out that my streaming service was collecting and analyzing my customer data for any purpose other than billing, I’d drop them in a heartbeat and move on to the next content provider. And they can take their movie recommendations and drive off a fiscal cliff with them.

  3. Their steaming catalog, for what I’m sure are valid business realities, is about as deep as a kiddie pool. There’s no content to personalize; and their original content is meant to appeal as broadly as possible. It’s a different business than the one which launched the Netflix prize.

  4. Modelling are tremendous valuable when events are stable, ceteris paribus. In those cases understanding patterns is key to predict and antecipate . And I can atest it does in many industries. Unthinkable to abandon models at profit in many segments. But models are simply models. They look to the past (or near present), correlate variable, test algortihms and then you have more chance to be succesfull. Not more than that. But better then coins….

  5. Netflix use personalisation extensively. They did implement a aspects from the netflix prize, they just didn’t use the combined final model (Which was an ensemble of the models they did implement).

    Take a look at any number of presentations at
    http://www.slideshare.net/xamat

    to get an understanding of the approaches they use to personalise the service.

  6. thank you for a thought provoking article.

    some thoughts:

    Don peppers and Martha Rogers wrote about a process; identification, differenciation, interaction and optimization. Its about creating a relationship.
    SETH Godin wrote about creating the relationship and permission.

    personalization is is more than making predictions based on previous behavior. that’s only one aspect of the relationship.

    novelty aspect is great in creating initial interest. But could it sustain the challenge of time without the relationship ?
    true novelty can be expensive (although we can create the illusion of novelty by working working on presentation /packaging and distribution).

    in some cases, the novelty and relevance aspects can be combined.

    I wounded if the novelty factor is as important in a B2B context.

    thanks again

  7. Good article – but i dont think that Netflix and their model can be the correct model for personalisation. What netflix have effectively done is bring the old video store feel to their online service. If you remember going to a blockbuster in the 90s you would be inundated with utter garbage films. The new releases which were most in demand would be the films with most stock and visibility in the store.

    I still dont believe that the word personalistion exists in digital marketing – to be honest the concept itself is too board. What is referred to as personalistion in many cases marketeers simply mean optimization.

  8. But what makes you think ran away from personnalization in the first place? Do you have insider knowledge on their next UI rewamp?

    I don’t have a netflix account myself but on the screenshot you give the first row, that is the most important part of the user interface is precisely a set of personnalized recommendations. Netflix also introduced user profiles last year to better address the multi-user account issue by providing personnalizations tailored for each individual user sharing the account:

    http://techcrunch.com/2013/08/01/netflix-user-profiles/

    If netflix had decided to “walk away from personnalization” the first row would likely be “Popular on Netflix” or “Newly released” lines and then you would have a prominent search engine or category browsing-based user interface. On the screenshot you give this is simply not the case.

    Furthermore Netflix is very very active on the recommender / personnalization research front. Just read blog posts by Xavier Amatriain and others at:

    http://techblog.netflix.com/search/label/personalization
    and
    http://technocalifornia.blogspot.fr/

    I believe that if Netflix is pushing the personnalized UI so hard over a generic search-based or browse-based UI it most probably means that personnalization works as a mean to increase their users’ engagement and retention. But obviously no-one outside of netflix has access to this kind of sensitive business metrics.

  9. I’m sure you’re not a technical person, because you misunderstood a couple of things:

    “What happened? Netflix at the time said the technical work of implementing the new personalization would be too costly for the anticipated return.”

    It was not about “technical work” in the sense of time to code the solution. It was about the time the solution took to learn and process recommendations. What Netflix said was that the winner solution was slower/less efficient than what they had, and the increase in recommendation quality didn’t justify spending money on more machines.

    “since bundling a new math model into a computer system surely doesn’t cost more than a bit of coding.”
    A math model is not just code, it has a cost. How fast it runs? How many machines I must have to run this model for 1 million users. It varies between algorithms/solutions/models. That was the case of Netflix decision.

    Overall, Netflix still has a lot of recommendation

  10. The main point is well-taken: algorithmic personalization is tough. Yep, it’s tough, and largely because it’s not currently easy for the algorithm to read the context that wraps a given user’s experience at any point in time. Are they shopping for themselves? Are they viewing with their family or by themselves. You note that, and my guess is that these will be problems that companies like Netflix, Amazon and the like will *have* to address.

    The novelty point is a good one, but it is not antithetical to personalization. They are, in fact, linked. As the good folks at NBC used to say during the rerun season, “if you have seen it, it’s new to you.” Personalization is designed to surface that which we have not yet seen, but which certain patterns suggest we should.

    Finally, the most important versions of personalization on the market right now are less driven by algorithm and more driven by interest graphs. Google’s “People Also Searched For” is one example. Amazon’s “Customers Who Bought This Item Also Bought” is another. This is not the same personalization as simply saying “show me what I want to see”, but it is nevertheless personalization that is simply coached by us offering a hint or seed from which the system can then look at others’ interests and map them back to mine for recommendations that are usually quite valuable.

Leave a Reply to Gideon Rosenblatt Cancel reply

Your email address will not be published. Required fields are marked *

* Copy This Password *

* Type Or Paste Password Here *