In Depth
Web Video Hurdle: Good Recommendations
Online video sites want to get to know you better.
The Sony-owned Webvideo destination Crackle is developing tools to offer personalized video recommendations to site visitors based on their past viewing history. NBC-Fox venture Hulu introduced a similar feature earlier this month.
They join video sites like Veoh and TheWB.com that offer varying degrees of viewing suggestions for visitors.
Tailored tips have worked well for Amazon and Netflix. On a video site, recommendations can pay off by keeping viewers glued to the screen longer, exposing them to more advertising and deepening loyalty. If Crackle or Hulu can mimic a good waiter, offering up the TV shows, movies and clips you like, you’re more apt to come back.
Offering the right recommendations is a challenge for video sites, however, because a Web site needs lots of data on user likes and dislikes to serve up quality suggestions. Amazon and Netflix work well because their sites attract large audiences. But cracking the recommendation code is more critical than ever for Web video sites as online viewing continues to swell. Internet users in the United States watched 13.5 billion videos online in October, a 45% jump from a year ago, according to comScore.
“Given all the choices, people are increasingly looking for guidance to find content and experiences that appeal to them,” said Eric Berger, senior VP of digital networks at Sony Pictures Television, which owns Crackle. “We’re building out the capability to recommend content to viewers based on what they are watching.”
A recommendation engine acts like a guide and can help users find videos that aren’t on the front page or featured section of a site. That’s helpful to Hulu, which now offers more than 30,000 videos from more than 130 programmers.
“Hulu has so much content that it's in a unique position to provide recommendations based on content type, right down to the episode of a particular show,” said James McQuivey, analyst with Forrester Research. “It's an area that can go awry, however, as when Hulu recommended an old ’70s movie called ‘Heroes’ to me because I watch the TV show ‘Heroes’ on Hulu all the time. They'll fix those obvious quirks, of course, but it does show how hard it is to connect people’s past behaviors to future desires.”
Hulu plans to roll out additional innovations in browse, search and recommendations, said Eric Feng, chief technology officer at the site.
Veoh has been offering recommendations since its 2004 launch; that has helped the site grow to more than 25 million unique users, said founder Dmitry Shapiro. If a user is logged in, Veoh can suggest videos based on past viewing habits across the site for movies, TV shows, Web series and user-generated clips.
Understanding viewing behavior such as what a visitor searches for, watches and then seeks out is critical to serving up an effective recommendation, said Adam Singolda, CEO of Taboola, a startup technology firm that provides video recommendation tools to sites like 5min.com and Aniboom.
“You need to study the way they behave in real time to recommend the next best video,” he said. “If they like short clips, you want to recommend shorter clips to them.”
This behavioral data works best when combined with intelligence about the actual content of a video, said Ben Weinberger, CEO of online video technology firm Digitalsmiths, which is refining the recommendation capabilities of its publishing tools for media companies.
“You want to learn if people are interested in comedy or comedy with a family theme, and then you can recommend videos like that,” he said.
Leave a comment
Comments 2
Phoebe
Interesting analysis of the video recommendations market today. The cold start problem - needing plenty of user data to generate good recommendations - relates to collaborative filtering methods that generate recommendations based on correlations among users. Recommendations based on content analysis (identifying the content elements that each user enjoys - as Pandora does with music) don't generally have the cold start problem.
Guest
This is not entirely true by definition – i'll explain:
It is true that Pandora has SGC (Studio Generated Content), thus from day 1 knows something about the content they need to recommend. Btw – so as Netflix, Amazon and so forth.
Should Pandora decide to take collaborative approach (to create interesting correlation between contents even - regardless to users at that point) – they will hit the same "cold start problem" situation you are referring to and will need to deal with it somehow. In essence, to "automatically" understand that A (content about A) is likely to be watched with B (content about B) (probability of y% for instance) an "old school" collaborative filtering would need some way to deal with the lack of data if both A and B were just "uploaded to the site".
Video recommendation has that problem for instance (i agree with you btw that video is far more complex because there is no pre-given knowledge about the video itself, like Pandora has).
2cents...