Do We Really Want Fine-Grained News Personalization?

I had some thoughts about personalization today, specifically personalization of news. Superficial research revealed the excellent Personalized News: A Market Overview from 2006 at RWW. More digging revealed the perceived difficulties of the approach: as (1) the model of personal interestingness and (2) the computation required. The market appears to be dominated by products that filter in two similar but subtly different ways: aggregates of your feeds highlighting the specific articles you are likely to read, and filtering the popular articles in the sphere and highlighting those that may interest you. There are standard claims that personalized news is implicitly important and desirable, and that personalization is the next big thing (web3.0).

I am not so sure that we as users want personalized news, at least not at super fine granularity. The obvious example is the case where personalized news works perfectly, and each morning I read only those articles produced in the last 24 hours that I am totally interested in. The effect is that I loose the shared experience of discussing news with colleagues and friends. I suspect the example highlights the cost/benefit of personalizing at the special interest group level, like 'technology', 'science', etc. More realistically, my interests would only ever be that specialized if I was working on project, or research topic (which is all the time). Therefore, in such circumstances I sacrifice shared experience (reading the paper, watching TV) for domain knowledge and resultant productivity.

Granular recommendation over RSS may be desirable presuming the model can track crazy human things like mood. This highlights the importance of the chosen model, and in particular the indicators used. For example, my mood suggests the cutoff on the type of articles I want to read and time I want to spend reading. A similar effect occurs on popular news items, for example, the articles I choose to read on reddit or digg. Another important part of the equation obviously is discovery, typically addressed along the Amazon-like lines of people who read X also read Y. End of the day, these points highlight the previously made point: "the model is hard". Google may have it right, abstracting the problem of personalized news to that of personalized search (specific interest).

I get digg, it is all about 'powered by the masses', but one thing that bothers me is reddit, which was designed as a news recommendation site and uses the digg (powered by the masses) as its front page (hot news). Why? I suspect that this is what people want to use, people are looking for the popular items that everyone is reading so they participated in the shared experience either directly (comment), or later on (latent acquired opinions). Additional evidence is Findory which was all about per-user personalization and fell by the way side at the end of 2007, after about 4 years with a modest user base. This makes me thing that either it is not a feature users want, or the difficulty of the model has not been addressed in any significant way as of yet to engender adoption. I don't think the latter is the case, because as with search, something is better than nothing. For example a crappy Altavista search was still better (adopted and popular) than not searching. This has not been tested (by me), specifically, how successful are the current crop? Is all the traffic still funneled to google reader and social news sites?

Chatting with D did provide an interesting point which I had neglected, which is the modeling of the easier problem of negative preferences (like negative databases). Such an approach could be used broadly to define special interest groups, or specifically to tune those posts of interest from an aggregate of feeds. Basically, the same as what people are doing now, although exploiting the generally stronger notions of dislike rather than the fuzzier notions of like.