Better Video Recommendations Require AI and the Smart Use of Data
As video providers expand their content libraries, their challenge is making sure viewers find the programming that appeals to them. Many services take immense pride in their complex content recommendation algorithms, but most recommendation tools today are collaborative filtering algorithms that recommend something based solely on an association to something the viewer watched in the past.
These learning algorithms are based on the relationship between things, which is probably why they’re not that successful. A 2018 study found that only 29% of viewers actually watch something recommended by a VOD service. For VOD providers to rise above the competition and connect with viewers, they’ll need to go beyond basic content recommendation and capture the larger opportunities in personalized experiences, especially in the mobile environment.
Netflix identified the importance of recommendations in driving user engagement back in 2006 when it created the “Netflix Prize,” a contest with a $1 million award available to anyone who could improve its movie recommendation accuracy by 10% or more. Working with a limited data set, it took three years until a team of engineers claimed the prize, but even then the new algorithm only just passed the 10% requirement.
A truly good recommendation engine—and a better user experience—relies on real artificial intelligence, which necessitates more data, from multiple locations. It goes beyond relatively simple “users who liked this also liked this” algorithms and stretches into “deep data.” This kind of intelligence comes at a cost, however. The second iteration of Netflix’s competition was supposed to use a much deeper dataset, with more than 100 million data points, including information about renters' ages, genders, ZIP codes, genre ratings, and previously chosen movies (all anonymized). But, the contest never got started as a multi-million dollar privacy lawsuit was filed after researchers argued that it was possible to “de-anonymize” the data. Privacy concerns are the largest challenge to the development of revolutionary recommendations, since any significant effort to innovate in this space will require test data sets that are closely guarded and personal. Innovators, therefore, must come to the table with a partner that has the kind of deep data that's needed for processing.
Integrations with social channels may be low-hanging fruit here, offering a much more complete vision of who a viewer really is and what they might want to watch. Linking VOD accounts to Twitter, Facebook, Instagram, and Pinterest profiles is a simple way to introduce more data points and make a huge step toward a more personalized experience, and a user’s likes and Twitter followers can paint an interesting picture of what a viewer might want to watch while using information that’s somewhat less-sensitive to today’s privacy issues.
Each of these social networks offers exponentially more data on a viewers’ interests beyond what they’ve watched before. Content recommendations can be based on actors they follow on Twitter or Instagram. As Netflix and others move into HGTV-like reality programming, Pinterest becomes a natural partner for pairing interest data with recommendations. Video providers can also simply make recommendations based on what a viewer’s friends have watched. This isn’t that far off from the associative models they use today, but it creates new opportunities for viewers to interact and bond around the content. By analyzing the data across social connections, providers can identify overlapping tastes among friend groups and possibly even promote viewing parties or spark ongoing online conversations.
Taking integrations a step further, services could identify characteristics of a user’s mobile device to provide additional context. A profile of which apps are installed on a device could inform what kind of content the owner might watch. In an extreme example, services could link directly to subscribers’ email accounts and calendars to mine personalization data there. If a service learns that a subscriber is buying hunting supplies, it can recommend a show like Discovery’s Raised Hunting. If Amazon isn’t already doing this with purchase data from across its site, then it’s missing the boat. If a Prime customer frequently buys diapers, they may be interested in children’s programming.
Of course, this isn’t the kind of service that is going to appeal to every subscriber, and strict privacy guidelines must be adhered to. Still, some portion of the subscriber base may be willing to grant read-only access to their online data footprint for a better experience, which ultimately helps the VOD service capture more of the viewer’s attention. With video providers more focused on engagement than ever before, it doesn’t hurt to ask.
In addition to deeper information about a user’s inferred interests, additional innovations could be made in how content is associated. Pandora made this intuitive leap with music, although its approach to solving the “music fingerprinting” problem was highly labor-intensive. The Pandora music service was founded on the initial principle that it’s possible to recommend music that someone wants to hear based on analyzing music someone has indicated they like. Pandora invested in musicologists, many with advanced degrees in music, to review just under 2 million tracks and fill out a matrix of up to 450 characteristics about each song. These expert associations go much deeper than typical collaborative filters in trying to identify similar content, but anyone who’s used Pandora will tell you existing recommendations based on this approach are not perfect. This type of approach, if applied to video, could yield some interesting results, but only if the video “genome” could be automatically generated, which implies sophisticated AI that can predict anything from thematic subjects to emotions a video might evoke.
Even without delving into deep data, it’s possible for video providers start personalizing content recommendations. They can start by simply slicing their own viewership data to better understand viewing patterns, and make better recommendations off of existing knowledge. For example, the issue with associative recommendations is they don’t currently discern between viewers. If your kids watch a lot of children’s content in the morning, that influences the algorithm when you settle in to watch something at 10 PM. By simply identifying user patterns and preferences based on the time of day, services can reduce some of the friction in the user experience and provide recommendations that the current viewer is more likely to engage with.
Once a service provider has refined the ways it makes recommendations, there’s the matter of delivering them. It’s easy to joke about endlessly scrolling through VOD menus looking for something to watch. What if the service downloaded the best content recommendation directly to your device in advance? Amazon has experimented with this idea, automatically downloading video recommendations onto users’ Kindle Fire devices via a service it calls On Deck. With growing mobile video viewership, it’s easy to see the motivation, but the number of forum posts asking how to turn off On Deck downloads show that this misstep failed to think of the user experience. Amazon crossed a line in the consumers' minds, turning recommendations into a creepy experience, akin to Apple and U2’s infamous forced album download blunder.
Just like with gathering the data to improve AI, the best method for delivering a new recommendations experience is permissions-based, with a clear opt-in to download content. Netflix’s smart download service, now available on Android, Windows 10, and iOS devices, is a step in the right direction, although it currently only downloads the next episode of a series the user is watching, rather than a new recommendation. Still, asking users if they’d like to download content—the same way they should be asked about access to social media or email data—opens up the opportunity for more customized recommendations, which should capture more of the engagement the service providers are after.
[This is a vendor-contributed article from Penthera. Streaming Media accepts articles from vendors based solely on their value to our readers.]
Brian Kline, CEO of Penthera, discusses how his company is maximizing KPIs for OTT publishing through the implementation of their products, such as their innovative new software PlayAssure
Network clips that display tune-in information are automatically suppressed by Facebook's AI, says BET, forcing the network to spend more on promotion.
The hottest driver off OTT growth, according to Amazon's Richard Au and Dailymotion's Anthony Layser, is customized, user-specific discovery algorithms, which they explain how to leverage in this clip from Streaming Media West 2016.
When consumers have access to an almost unlimited supply of premium content, surfacing programs they'll enjoy becomes a major challenge.
Get more video views while boosting engagement and advertising. That's the promise of JW Player's data-driven recommendation engine.
The algorithm wizards at Netflix pull back the curtain and show how they use worldwide data to improve local and personal recommendations.
ThinkAnalytics debuts ThinkBigData, a big data platform for understanding and gaining insights from billions of monthly customer data points.