Going social: Recommendations engines need to factor in consumer reviews November 28, 2011Posted by David Card in Uncategorized.
Tags: Amazon Inc., collaborative filtering, e-commerce recommendations engines, IHeartRadio, interest graph, liked product, media data, media recommendations, Online music stores, recommendations, reviews and recommendations, social commerce, social data, taste graph, Twitter Inc.
Last week eBay announced it was buying Hunch, reportedly for $80 million. Even though it has its own recommendations system based on searches and popular items, eBay grabbed Hunch’s recommendations engine. The acquisition illustrates the trend of adding social media to the data-powered recommendations credited with increasing Amazon’s sales by 25 percent. But get ready, because the next stage of commerce recommendations will further mix in human-powered consumer reviews.
Pioneered by Amazon, the first wave of e-commerce recommendations engines were based primarily on collaborative filtering. That is, they compared a lot of data: “Customers who bought product x also liked product y.” As with most big data analysis, recommendations are not only about pattern matching but also about which patterns best predict future purchases. And like all pattern matching, the company with the most data usually wins. Social media, via open APIs from Facebook and Twitter, bring new data for analysis and lower the cost of entry for recommendations engines.
Game-changing social data from APIs
Hunch creates a “taste graph” of a user’s interests based on analyzing available social media data from Facebook, Twitter and others that is fine-tuned by user responses to a Q&A process on the Hunch site. That’s why it could deliver useful results without suffering from a “cold start,” or lack of enough data to be predictive.
Effective e-commerce and media recommendations will mix in big-data analysis of different information sources including product factors, observed purchase behavior, and social and interest graph data from APIs. They will create algorithm-driven engines from that data and present it alongside customer reviews that themselves can be filtered and analyzed by adding structure through categories, comparisons and ratings schemes the way Amazon and Best Buy do.
Another recommendations pioneer, Pandora creates personalized radio stations based on factor-based analysis somewhat like Amazon. Pandora programs based on mapping and matching song and artist characteristics, along with a user’s likes or dislikes. Its effectiveness makes it the leader in online radio, with 40 million active users and $75 million in third-quarter revenues. Lately, Pandora has been scrambling to build out social features to enhance its website and apps and get more inputs.
A startup with an MIT Media Lab background, The Echo Nest has begun licensing its music recommendations engine to online music programmers like KCRW and Clear Channel’s IHeartRadio. The Echo Nest automates its song-characteristic analysis by running each track through audio analysis. It also scans online blogs and music pubs for information that informs its audio analysis. This makes its approach more scalable — and cheaper — than Pandora’s engine.
A blended approach in the future
The established engine technologies still work, and social media data lets new players build them effectively. Last year, Amazon started experimenting with available social data. When connected via Facebook, Amazon remembers a user’s friends’ birthdays, suggests items popular among friends and makes recommendations based on favorites the user has listed explicitly on his Facebook profile. But so far, all of those results are on a separate, “beta” page of recommendations that Amazon links to from its personal recommendations page. They’re not mingled with Amazon’s traditional recommendations and user reviews on product pages — yet. Likewise, as it learns what social data is most predictive, I expect Amazon will incorporate the social signals it gathers from Facebook connections into its recommendations-ranking algorithms, the way Microsoft uses Twitter and Facebook in search.
Successful recommendations will take this approach of blending social and data. At GigaOM’s RoadMap conference, Wal-Mart sketched how its social commerce strategy, driven by its Kosmix acquisition, would focus on search, recommendations and local, in-store context rather than stores on Facebook. Wal-Mart is one retailer with serious data smarts. This is just one example of how, for commerce, it’s wisest to think of social media as data sources rather than shopping hubs.