syre / inferring-social-networks-from-geographic-coincidences Goto Github PK

Friends should perhaps be defined based on percentage of cooccurrences rather than just a fixed value

Statistics over co-occurrences

x-axis is for example [0-50000,50000-100000,300000-350000]

Advantages/Disadvantages of using only 2 decimal degrees precision (ease of computation, might include a spatial bin which covers a lot more area than necessary) vs full precision

Create histogram/boxplot for features which we want to investigate, suggestions welcome @handiandi !

Start the reverse geocoding script "geocoder_script.py"

Right now we classify a friend based on if they have more cooccurrences than 5 and that they are alone in those cooccurrences, is this approach fine? (it doesn't account for people with lots of cooccurrences)

right now a timebin size is defined by the range (start_time - bin_size/2) to (end_time + bin_size/2) which leads to varying timebin sizes

Review code for co-occurrences

Right now cooccurrences can be counted several times for each spatial and timebin, they should only be counted once (check the crandall paper!)

Histogram for co-occurrences binned by day, where do co-occurrences happen most frequently?

Right now in our code one user can have multiple nationalities, to make spatial binning for a single country possible we need to assign a single nationality for a user (preferably based on which country is most occurring in his location updates)

Is this an issue?

Filter data so we only get data where 0 <= accuracy <= y, where y=100.000 (mm)

Maybe y=y+20.000

The corr feature cannot be implemented right now because of the number of possible spatial locations is extremely large (60+ billion), this can be alleviated with sparse arrays but correlation cannot be computed to our knowledge with sparse arrays!

When making histograms allow to set a self-defined max value (steps will then count up to that value), and the last step will be a bin of > max val