In the old days, futuristic beeping devices allowed James Bond to track the villain and come in just in time to save the world (and flee with the girl kept prisoner, of course!). Nowadays things have grown a little bit more complex. Position alone is not enough: the whole behavior of any object, moving or otherwise, must be taken into account when tracking it. Adding information on the activities of an object to its position gives us something called an activity trajectory. Of course, nowadays we have sources of information James Bond could never dream of (not even with the help of Q): social networks.
The problem here is that social network data are usually unstructured; more specifically, problems arise when trying to index and query these data in order to extract knowledge from them. Extensive research over the past years has built solid results on positional information; current methods basically check each trajectory, defined as a succession of positions, compute how closely it matches the query, and return the top results. The authors of this paper move beyond positional information alone and propose top-k spatial keyword (TkSK), a method to query activity trajectories that can reveal preferred combinations of data from which to infer recurrent patterns. In a nutshell, it defines activity trajectories as finite sequences of timestamped locations to which some keywords are added. The authors call this process the baseline algorithm.
As noted previously, data are acquired through social networks, namely Foursquare, Facebook, Bikely, and Flickr. What the method proposes to its final users are several location sequences, based not only on geographical proximity, but also on choices made by people in the past. From technical and theoretical points of view (which matter most to an academic audience), the method is composed of a similarity function and an indexing method that together improve data query efficiency over existing methods.
The paper first defines the problem formally, and then presents all of the components described above. In its final section, it gives some experimental results and points to directions for future work.
In particular, experimental results show how TkSK performance varies when playing with several parameters, including the number of top results desired, the number of keywords and parameters used, the number of trajectories analyzed, and the time window these trajectories extend over. Of course, as the number and type of parameters grow, so does the load on computational resources available, as well as the overall accuracy of TkSK. As for future work, improved performance will be achieved by adding the ability to modify keyword ordering in the queries, by using vector modeling to compute text relevance within the queries, and by linking together successions of saved queries.
Of course, in this paper, TkSK is just an algorithm, defined in mathematical terms and as pseudocode, so the preferred audience for this paper would certainly be academic. If that sequence of keyworded locations could someday be turned into a tangible product, laypeople (or even James Bond) would find interest in this method.
