I have geolocational data(coordinates and times with device id), I can bucket this down using say 5m by 5m squares to represent a vertice on a graph. Then following the device id and creating edges chronologically creates a walk(or a graph). I also have a label for each walk, either a True or a False on whether they visited a particular location or not.

I want to use a supervised machine learning method to use this data to predict whether future walks would be a True or a False. One idea that I'm currently exploring is to generate a graph kernel with unit edge weight and try logistic regression or support vector machines on that.

Does anyone else have any ideas on other machine learning methods that I could explore? Is there a way of leveraging the time-series data such that I don't have to even create the graph? I should also mention that I have data on walks that I don't know if they correspond to a True or a False as well, so a semi-supervised method would be welcome as well. Thanks

  • $\begingroup$ Do you have an actual graph (edges and nodes) for the paths that these points lay upon? Is the actual point to have the machine learning algorithm infer the geographic location of the particular location that is potentially visited, so that any walk through a proximity around that point triggers a prediction of True? $\endgroup$ – John Yetter Jul 6 '16 at 12:58
  • $\begingroup$ @JohnYetter No, all I have are GPS co-ordinates, I plan on creating a graph as a way of mitigating the fact that I don't know how to handle the time-series data. I was thinking of creating a grid of 5m by 5m squares and a measurement in one box followed by another measurement in another box would be an edge connecting two vertices. I'm not sure how two measurements in the same 'box' would be represented.As for your second question, yes that is roughly true but my goal is to predict based on movement behavior of visitors(as opposed to someone who goes past in a straight line). $\endgroup$ – Jimbo Jul 6 '16 at 13:22
  • $\begingroup$ I don't think that you need to deal with the data as time series, unless the time of day (or within the week) that they approached the spot is important. The main thing is that they approached that geographic location and stopped there. Does it matter where they approached from? That is, does order matter, or just proximity? $\endgroup$ – John Yetter Jul 6 '16 at 16:04
  • $\begingroup$ Order does matter. Proximity is the definitely the easiest way of inferring a "stop", I want to see if I can use the order/or the sequential nature of the measurements as well. $\endgroup$ – Jimbo Jul 6 '16 at 16:22

//// Can't comment because of reputation so adding an answer ///

Please refer to below paper -


The main idea of the methodology is to compute probability of each possible path & then model it

Alternately you can use Markov chains if the deviceIds can only move between set destinations (like bus stops or tourist points) -


| cite | improve this answer | |
  • $\begingroup$ I think if you expanded each of these options into a slightly longer description of what's going on "under the hood" you would have a very strong answer that could dispense with the "not enough reputation to comment" disclaimer. $\endgroup$ – Sycorax Jul 11 '16 at 19:35
  • $\begingroup$ Hey I accepted because graphical models definitely seem to be the direction to explore. I'm currently looking at Conditional Random Fields as well and trying to think of how to work that with my data. $\endgroup$ – Jimbo Jul 12 '16 at 13:30
  • $\begingroup$ @GeneralAbrial - I'll keep that in mind for next answers :). I didn't want to sound naive since I don't have a proper grip on concepts of conditional random fields. Jimbo - If the method helps/doesn't please put a short description here. Would love to know the application of this to your problem ! $\endgroup$ – wololo Jul 13 '16 at 7:24

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.