Uncanny Valley in Customer Profiling and Recommendation Systems?

In the Data Mining course I’ve been doing we’ve been discussing our position as customers whose data is used to model us and target services or products based on that analysis. One of the examples from another student was of their mother being annoyed by gambling-related books after purchasing a book on Amazon. It very much seemed that some of the key customers for horse-related books are gamblers – presumably those who bet on the horses and spend time understanding the runners properly. However, as a non-gambler she was not happy to be targetted by those books, so it was a situation where their previous purchase put them in a cluster which was not representative of their own interests and created a negative response rather than a positive one.

It got me thinking – is there an ‘Uncanny Valley’ effect when profiling customers and adapting your offerings to that profile? In case you’re not familiar, Uncanny Valley is a term used with respect to visual effects and robotics, where when representing or imitating human shapes and features there is generally an increasing likeability up to a point, and then a dramatic drop-off and rejection, before the result becomes more likeable again. How strong this effect is seems to vary culturally, with Japan seeming more tolerant of humanoid robots than other cultures, for example, but discussion of films like Beowulf, Polar Express and Final Fantasy touch upon this subject. The prevalent theory seems to be that the robots or images look more like corpses than living humans, because they lack certain detail or movement that would make them seem more alive or real, and they then seem very creepy indeed. Once the images become sufficiently realistic they no longer look creepy and we respond to them more like other people and like them more again.

With respect to profiling and targeting, I’m wondering you have a similar valley effect. Your baseline is obviously going to be having no customer profiling at all – everything is just aimed at the population as a whole. You would expect that if there is some non-obvious profiling – with very broad groups – this is likely to make the offering without making the customer feel like their privacy has been violated. While the profiling is not obvious to the customer, more accurate offerings can be provided and the appeal will increase. At some point, however, the profiling is doing to start getting more specific, and then you’ll start to notice more products which are similar to those you’ve already looked at, or which you don’t think would be that common, but which are being listed quite frequently.

At this point, if the recommendations are good, then this may encourage further sales. However, since the profiling is imperfect then there will be some which are incorrect, possibly completely wrong. In the case of a non-gambler being offered lots of gambling books, this may be borderline offensive, just as being over-targeted for debt assistance services or the like would offend others. Another source of problems is when a family ends up profiled together from sharing a machine or account, and you find yourself bombarded with adverts for items which would be relevant for a relative rather than yourself. This is where the targetting becomes obvious but the accuracy is not yet good enough to be a pleasure, and where the greatest chance of offence lies. This is what I’d consider the ‘uncanny valley’ of profiling. Whether you liken it to cyber-stalking or ‘Big Brother’, it seems intrusive on your privacy as well as being jarring because it presents an incorrect profile to you when it makes a mistake.

At the other side of this valley the number of mistakes decreases, and you have less potential for causing offence by making incorrect assumptions about your interests. Instead, as the accuracy improves it will naturally offer more appealing offerings and making finding what you’re after. I’d liken it to a personal assistant, who gets to know your likes and dislikes. The dislikes are possibly the key element, as they don’t present you with things you don’t want or even object to, because they know you well enough, and don’t trigger the negative responses that incorrect targeting can. Highly accurate profiling allows the right products to be presented at the right times to maximise your uptake and engage you as much as possible as a customer. Even here there is a balance in that someone can feel led into temptation and have too much sold to them, perhaps, but since this is another part of a full profile a perfect classification would make as many sales as it could without damaging the relationship by pushing too much.

I don’t have the data to analyse to verify this trough of rejection at the point of increasingly obvious but visibly inaccurate profiling, but anecdotal scenarios fit with it and logically it makes a degree of sense. One question would be how to verify the concept, or alternatively, if you assume that it is likely to be correct, how can you best avoid alienating customers at this point? I don’t have any easy answers for the former, but in the case of the latter at least you can perhaps limit the amount of targeting to a threshold to minimise hitting the uncanny valley effect, while monitoring further engagement to get a clearer idea about the customer. If you’re very confident that the profile is correct then you can fully use the profile you’ve created. You just have to limit profiling to avoid being too obvious until you’re sure. In the case of books, In the original example, for someone who has bought a book on horses, further books about horses may be appealing and directly related to the previous purchase, but extrapolating to say someone is likely to be interested in other animals a bit less sure, and ‘related subjects’ like gambling much less certain. Low response rates to all suggestions may make confidence levels hard to use. Perhaps some kind of negative scenario analysis would be required – to identify where promoting something results in a decreased engagement from the customer, and profiling the risk of customers falling in these groups and addressing that instead.

It’s worth some more thought, though, as many now share a great many intimate details about their lives quite freely, but can still feel pressured or invaded by unwanted and inadequate targeting.