A couple interesting if somewhat self-serving articles from ChoiceStream recently. ChoiceStream, among other things, is the company behind MyBestBets, a fun recommendation service for TV programs.
The first article claims that "more than 80 percent of online consumers are interested in personalization" and that "56 percent of respondents [are] willing to provide demographic data in exchange for personalized content." Good news for personalization solutions that rely on filling out lengthy profiles, but the conclusions are directly from a ChoiceStream survey, perhaps not the most objective source.
The second article is by the CEO of ChoiceStream, Steve Johnson. Steve claims we are experiencing a "personalization renaissance" where companies look to personalization to deal with an "unfiltered flood of information" and focus on satisfying existing customers in a more mature online marketplace. The article makes some interesting points, but ends with a bizarre claim that "Attributized Bayesian Choice Modeling" is the obvious solution to every personalization issue. Coincidently, ChoiceStream's product just happens to use this technique. While I don't have anything against this particular technique -- it's a content-based approach that might work well for some problems -- what I really object to is the (biased) claim that it's the obvious solution to all your personalization needs.
Personalization is hard, folks. Different approaches have different characteristics on different types of data. Some have higher predictive accuracy in some cases, some have better scalability. Not only do you have to pick the right approach for your problem and your data, but also the approach that works best may change over time as your data changes (especially as your user base grows). There aren't easy solutions. But, if you do put in the effort, personalization can be a powerful and sustainable advantage over your competition; no product your competitors buy off the shelf will replicate it.
Subscribe to:
Post Comments (Atom)
2 comments:
Hi, Daniel. Glad to hear you're enjoying Findory News!
To be clear, my objection is to setting up a false dichotomy between CF and ABCM when there are a wide variety of techniques to apply to this particular problem. Furthermore, while I have no a priori bias against ABCM, the claim that ABCM is better than the many other approaches, including forms of CF, on a broad class of problems is not well substantiated.
As the last paragraph of my post should make clear, the real issue here is that the algorithm used needs to match the data. As the data changes over time, different techniques may start to perform better. If personalization is to be a competitive advantage for a firm, it can't be an off-the-shelf product that the competition can also buy; it needs to be a custom solution tailored to the company's data and customers.
Post a Comment