When it comes to social search and social recommendations, there is a lot of hype around the concept, but given that a user’s social graph is – almost by default – limited, there are major gaps in both accuracy and coverage when it comes to putting this concept into reality. While Google +1 and Bing’s implementation of Facebook ‘like’ data are trying to find ways around this, Microsoft researcher Mohammad Raza argues (PDF) that we need a smarter recommendation system that is based on the idea that “your friends know you and can best predict your taste” and that social search can be greatly improved upon with the help of prediction extraction.
Note: you can download Raza’s paper here.
Flaws in Today’s Social Search Schemes
To understand why this matters, let’s look at how most of today’s social search and recommendation systems work. As Raza points out, the two main ideas behind social search today are that “your friends are like you” and “people we agree on certain things also agree on others.” In reality, though, individuals and communities are far more complex than this. Users may agree with friends on some things (pizza), but disagree on others (politics). While Razza doesn’t go into this, it’s also worth noting that the idea of “friendship” on social networks has become so diluted that many of you “friends” today have little to none in common with you.
Then, there’s the problem of coverage. Users generally only talk about and rate items when they have a strong positive or negative opinion about something. “Part of the difficulty,” writes Raza, “is to motivate people to give more feedback on more mundane items, or items that may be important to different people under different circumstances.”
How to Fix This?
Raza argues that we can past these problems by getting users’ opinions about others’ opinions. Even if your friends don’t agree with you about everything, they are, says Raza, “actually the people who know you best” (his emphasis). The idea the, is to “elicit predictions about the target user’s opinion of a certain item from the user’s friends who have experienced the item, and aggregate these predictions to construct an estimation of the target user’s opinion of the item before he has experienced it.”
Raza proposes to use a Facebook app that allows users to rate items they have experiences (movies, news, events, food, YouTube videos etc.). The unique twist here is that this app will also ask users to predict how one or more of their friends would rate this item.
Once a user then actually experiences this item (say a YouTube video) and rates it, this score will be used to train the algorithm and the software can learn which of your friends know you best and take their ratings into account when presenting you search results or other recommendations. Of course, the algorithm will also learn if your friends are good at predicting anybody’s reaction in general or if they are just good at predicting your reaction in certain areas.
Raza also proposes to push this system even further by allowing users to give reasons why they think a friend would like an item and elaborate on their opinions. Say your friend thinks you will like the movie “Inception” because it has Leonardo DiCaprio in it or because large parts of it play in Paris. The algorithm will then know that these are things you like (assuming your friend has been classified as trustworthy) and can tweak its recommendations accordingly.
For now, of course, there are no public implementations of this idea, but it does sound intriguing. In my experience, I find myself drawn more to purely algorithmic recommendation systems like my6sense and Zite than social apps like Flipboard and News.me because they have come to know my tastes better than my wide-flung group of friends on Facebook and Twitter. Chances are, though, if these social recommendation algorithms knew which friends to trust and who knows me best, this hybrid system that pulls in a far wider range of signals could present me with better recommendations than either system alone ever could.