Exporting the Amazon model for behavioral profiling
In the book Profiling Machines, new media theorist Greg Elmer looks at how consumers are tracked and how their data is fed back to them, affecting change in the spaces they navigate. His examples include recommendation systems like those on Amazon and TiVo. On either, your behavior or ‘movement’ is recorded and analyzed along with that of every other user, generating various profiles and thereby recommendations. I think that the key point is that these systems generate change in the user’s environment while promoting uniformity in users’ movement.
Below I wonder aloud, “What would be the effect of applying this model more generally?” (Note: 75% of readers ultimately choose Chuck Norris after viewing this item.)
The key words here are navigation, profiling and cybernetics. That is, we’re looking at ways to smooth communication between user and system, not entirely unlike Norbert Wiener smoothed the interaction between anti-aircraft guns and their users.
Lately I’ve been interested in how the Amazon model is being exported into other systems, including some with (I believe) real benefits. But first we need to list some of its key aspects:
- A closed system. Effective behavioral profiling relies on having all the data, so it is hard to imagine it working across TiVo and Amazon, for example.
- Uniquely identifable users. On Amazon, we can remain anonymous and still get personalized recommendations each time we visit, since the site uses a cookie to remember us. The point is that while we can remain anonymous we must be identifiable.
- Noise filtering. I’m only guessing here, but it would seem necessary to develop defences against possible noise – e.g. Amazon somehow filtering out the movement of webcrawlers, or the erratic movement from IP addresses that are shared by a number of people.
- Automated behavioral analysis and feedback. How complex can this become? I would assume that an understanding of dynamic systems or chaos theory would help illuminate this aspect. Remember that these systems not only have to account for a growing or changing set of users, but also changes in the range of products they offer. So it is logical to think that some kind of modeling is done in a process of tracking change as well as predicting it (e.g. how does the recommendation system affect, say, the quantities of stock).
- Data Retention. This one’s obvious.
The above is just a few things off the top of my head, but could be a starting point. Now for the fun part: how might this model serve other purposes?
- Personal Communications – we’re already used to spam filtering, whether it works well or not. What about an email provider that is also your secretary, sorting all your mail for you based on past behaviour (e.g. you always drop whatever you’re doing to read and reply to your best friend, so it notifies you immediately, but waits till you’re not busy to tell you about the latest email petition). There’s work being done for this, as reported in the New York Times. Perhaps such a system could be part of 4th Gen. mobile phones to sort all of your communications.
- Reshaping Physical Space – last week in class someone mentioned how tracking mobile phones is changing the way we deal with traffic jams (a big issue in holland), since, with the cooperation of the telcoms, we can ‘see’ them form in real time and even predict them. Tomtom, producer of mobile navigation devices, is looking into integrating this data to its own ‘recommendation system’ (Link is in Dutch). Note that while this is not a closed system, it will be once every car is fitted with a Tomtom or its equivalent. The next logical step would be to use such data in city planning, thereby feeding back our collective preferences into the space we navigate. Using behavioral profiling to better organize physical space is not new: one of the reasons supermarkets track consumer buying habits is so that they can rearrange the shelves at different locations and meet the ‘needs’ of a particular set of buyers. For an example visit a supermarket chain’s locations in high-income and low-income areas, and see if you can spot significant differences in pricing, arrangement, product line, etc (or just see a related article in the Economist).
- Reputation Models – This one is new to me. Mark Pesce, while he doesn’t give it a name, talks about improving on previous reputation models by analyzing behavior. He responds to the recent Craig’s List experiment (warning: sexually explicit), in which a user pretended to be a girl looking for sex and then published all the responses he got. Pesce argues that Craig’s List could track and profile users based on their navigation of the site and their established network:
Craigslist does have a login capability, so it can potentially record each of the interactions users have through the system. It could collect data about the quality of the trust interactions users experience on Craigslist, and use this information to annotate all of the postings on the system. In short, every posting on Craigslist could be accompanied by metadata which allows users to have some basic sense of the trustworthiness of the other participant in a given transaction. With each successive transaction, Craigslist could begin to model an emergent digital social network
Greg Elmer (2004) Profiling Machines
William Bogard (2000) Smoothing Machines