My friend Chris Done is an extremely talented and imaginative developer who as far as I can see has only one major flaw — his blog doesn’t allow comments. So I’m going to have to comment on his latest harebrained scheme (his words;-) on my blog, and hope that his can pick up trackbacks …

The idea is to build a movie/TV/[other] review site based on links of trust, in other words one where the reviews that are brought to your attention aren’t the same for everyone (like a newspaper’s review section) but are filtered according to your network of friends or trusted reviewers.

This is a fun idea for two reasons. First, it offers a simple, scalable means of filtering out reviews or reviewers that are bad, or consistently not in line with your taste. (Actually, I’m not sure this is always such a good thing. I once went to a play precisely because it had been slated by a reviewer who had previously slated another play that I loved.) It’s a much simpler problem computationally, and it can also be readily implemented as an unhosted web application, rather than on the basis of a massive centralized dataset.

Second, although the general assumption is that you’ll trust people whose tastes are similar to your own, a social network approach should help to ensure the presence of those weak ties that can be so important in ensuring diversity of recommendations. Typical collaborative filtering approaches based on similarity measures tend to eliminate the possibility of surprising or serendipitous items being recommended. But through our weaker social connections, we are more readily able to discover unexpected possibilities. Making use of stronger and weaker social connections would represent an interesting alternative approach to the diversity/accuracy dilemma with which recommender systems are often plagued, but which (shameful self-promotion here) it is possible to overcome.

Anyway, with that in mind, here are some questions to ponder.

• What does the network of reviewer trust look like? This immediately calls to mind Twitter more than Facebook — trust in particular skills isn’t necessarily reciprocal (e.g. I surely trust Chris’ software development skills far more than he trusts mine), so you’ll see followers more than friends, i.e. a directed network. In addition, evidence from Amazon suggests that the Pareto Principle is at work, with a fat-tailed and probably power-law distribution of review authorship. In other words, a small minority of very active reviewers are responsible for the majority of reviews written; most people don’t write anything.
• How can such a system be plugged into existing social networks? While most people aren’t writing in-depth reviews, most of us do at least occasionally use social networks to share recommendations of things we’ve liked or that we are approaching with anticipation. There’s a huge amount of data here waiting to be tapped. An ideal ‘trusted reviews’ system would be able to coordinate data sources from many different social networks, and to separate out and classify media recommendations from the rest of the traffic, even (especially?) the inadvertent ones. It might also provide an easy way to distinguish between which social contacts’ opinions you consider relevant, and which you want to ignore. Some friends I love dearly, but our film and music tastes just don’t overlap …
• Can it be integrated with existing review services? Many ‘proper’ reviews get written directly on e-commerce sites like Amazon, Netflix etc. The social review network could offer a means for people to post reviews across multiple sites, or to collate and broadcast in one place all their reviews written in multiple different locations. You might also want to consider how you handle reviews vs. ratings (if you want to handle the latter at all, as there are some interesting biases in the distribution of online ratings).

Happy coding, Chris, and thanks for provoking some (hopefully) interesting thoughts.

Update: Forgot to include one remark on the network of reviewer trust. There’s at least one such network (although it’s not a particularly social one) that could be analysed, the Amazon data on people-who-found-such-and-such-a-review useful. One could also look at Facebook ‘likes’ of media posted by users. Although these are different from the more explicit ‘network of trust’ that Chris proposes, they might offer useful clues as to its expected structure. Unfortunately in neither case does the data seem to be available … :-(

Update 2: Chris has explained why his blog doesn’t permit comments.