A study about mobile phone location data and recommendation systems:
People who played with location-based recommendation systems may have been confronted to a common issue: when you start using the application, you do not necessarily have a "location history" (no list of past "check-in" if we translate this in the Foursquare idiom), hence it's difficult to get relevant recommendations. This phenomenon has been called "the mobile cold-start problem" in this paper. This academic article written by Quercia et al. for the IEEE ICDM 2010 conference addresses this problem in the context of mobile recommendation systems, apps that can identify patterns in people’s movements in order to recommend events and services. The researchers investigated how social events can be recommended to a cold-start user based only on his home location. They conducted a quantitative study to investigate the relationship between preferences for social events and geography. They tested a different set of algorithms for recommending social events and evaluated their effectiveness.
Some excerpts of the results that caught my attention:
"In a situation of cold start (user preferences are unknown), recommending geographically close events produces the least effective recommendations, while the most effective recommendations are produced by recommending social events popular among residents of a specific area. (...) Interestingly, there are geographic areas that are more predictable than others, and this does not depend on the number of residents we consider in each area. We are trying to obtain sociodemographic data for Greater Boston to test whether sociodemographic factors such as income and inequality would explain those differences. If that would be the case, to produce effective recommendations, one would then need to complement real-time mobile data with historical sociodemographic data.""
And this bit about the data themselves is relevant too:
"To infer attendance at social events, one needs large sets of data of location estimations. Often such sets of data are not made available to the research community, mainly for privacy concerns. Such fears are not misplaced, but they gloss over the benefits of sharing data. That is why our research agenda has been focusing on situations in which people benefit from making part of their private, aggregate data available. This paper put forward the idea that, by sharing attendance at social events, people are able to receive quality recommendations of future events."
Why do I blog this? Working on the user experience of location-based services, I've always been curious about recommender systems and the problem designers face developing them. What's so fascinating is that they are based on basic and somewhat intuitive ideas about the way city-dwellers behave. Studies about their usage often reveal the complexity such systems.