idio Platform

Since it is what we have been doing for 2.5 years, and are now starting to do for others, I thought it would be worth getting some posts out about the various methods for content personalization. This first post looks at the major issue that any personalized system faces – the “cold start” problem. Until the ‘personalization engine’ knows enough about you to give you content, it cannot demonstrate its worth – therefore some method of jump-starting the process must be used.

The looong questionnaire:
Remember the days when you filled out four pages of detailed questions in order to get your free Hotmail account? Well, needless to say, those days are over. Users are very resistant to spending much time on setting up their login/profile to a new service, and regardless of how good your personalized content is, most users will never find out if you start them off with an interrogation. (Their are a few niche instances where this might work, however. For example, RealAge requires a detailed setup because it is clear that the ‘personalized health plan’ they give you could be detrimental if not customised correctly…)

The fun questionnaire:
Several services have made the ‘taste profile’ creation process a breeze by incorporating a game element. By this I mean that pages of questions are replaced by a fewer number of multiple-choice questions, usually represented by images. A great example is Magdentifier, which asks general questions about lifestyle taste, in order to recommend several print magazines for subscription. The danger of this approach, is that unless the options are chosen very carefully, the abstraction of questions away from the actual taxonomy of interests can lead to dubious results. There is a service called VisualDNA which builds an image-selection application for clients (a good live example is here – finding a similar celebrity and recommending fashion products).

Presets:
This approach divides the audience into segments (high-flying executive, fashion-conscious mum, etc), with each user selecting the best-fitting segment for them. The segment attributes are then used as a starting point for personalization for each user. A great example of this is Veedow, which asks you to select a lifestyle preset in order to recommend consumer products for purchase.

Attention-data import:
This is the holy grail for many in the Data Portability community, as it allows users to upload their taste profile to services as they see fit. idiomag, our personalized music magazine, is a great example of this (even if we do say so ourselves…). We allow users to create taste profiles by telling us their usernames on relevant social networks, or by directly uploading an APML file (attention profile markup language, the emerging standard for attention data portability). For example, by going to idiomag and typing in your last.fm username (or from one of 11 other services), you will immediately get a music magazine based on your listening habits over the past few months. Obviously this approach only works in verticals where users track taste, but the number of these are increasing rapidly, and off the top of my head I can think of social networks and services would would allow attention data import in the following verticals: music, movies, gaming, career, personal finance, lifestyle, gossip/celebs, fashion…

These are the major approaches to overcoming the “cold-start” problem. Of course, as soon as you have a reasonable idea of the user’s taste, the challenge is deliver the right content, and also allow further modification/evolution of the profile to increasingly provide a relevant experience.