gpsFExtraction.m

%% This file generates features for GPS - 
% Approach is as described in North Western study
%%
clc, clear all, close all
% Load gps data cell
load 'data\gps_feature.mat' % Last cell represents the user ids
%%
% Read data - change to loop later, for all users
for z = 1:(size(gpsData,2) - 1) % Last column has users id
    userData = gpsData{z}; 
    %%
    % Calculate how much time user spent on a perticular location
    % Traverse through all location samples 
    % clear timeSpent
    for i = 1:size(userData,1)
        if i == 1
            timeSpent(i,1) = 10; % At the start, time = 10 (sensing frequency) 
        else % See the previous long lat difference and current travel
            latDiff =abs(userData(i-1,2) - userData(i,2));
            longDiff = abs(userData(i-1,3) - userData(i,3));
            if (latDiff < 1 & longDiff < 1 & userData(i,4) == 0) % Combine case
                timeSpent(i,1) = (userData(i,1)/86400 + datenum(1970,1,1)) - (userData(i-1,1)/86400 + datenum(1970,1,1));
                timeSpent(i,1) = timeSpent(i,1) * 24 * 60; % to get minutes
            elseif (latDiff >= 1 | longDiff >= 1 | userData(i,4) == 1) % New ts
                timeSpent(i,1) = 10; % 10 minutes here
            else
                timeSpent(i,1) = userData(i,4);
            end
        end
    end
    userData(:, end+1) = timeSpent;
    %%
    % First feature - Location variance - that measures the variability in the
    % subject's GPS location - For stationary user only
    % Get the user data where state is stationary i.e 0 
    ind = find(userData(:,4) == 0); 
    userDataStat = userData(ind,:);
    % Calculate statistical variance of longitude and latitude
    varLat = var(userDataStat(:,2));
    varLong = var(userDataStat(:,3));
    locVar = (varLat + varLong); %Nautral Log
    gpsFeatures(z,1) = locVar;
    %% 
    % Entropy - to measure the variability of the time the subject spent at a
    % location cluster 
    % First identify clusters in location - use only stationary state data
    LongLatStat = [userDataStat(:,2) userDataStat(:,3)];
    
    kmax = 10;
    for k = 1: kmax
        [IDX, C, SUMD] = kmeans(LongLatStat, k); %returns the within-cluster sums of point-to-centroid distances in the K-by-1 vector sumD.
        Sum (k) = sum(SUMD);
    end
    Sum2 = Sum; Sum2(1) = [];Sum2(kmax) = Sum(kmax);
    Diff = Sum - Sum2; 
    Diff(kmax) = [];
    [Y, I] = min(Diff); % I is the proper number of k-means
    [idx, C] = kmeans(LongLatStat, I);
    gpsFeatures(z, 11) = I;
    
    
    
%     [idx,C] = kmeans(LongLatStat, 3);
    % Add cluster number to data
    clususerDataStat = [userDataStat(:,:) idx];
    % Percentage of time spent in the cluster - ClusterLoc/Overall time
    clus1Ind = find(clususerDataStat(:,end) == 1);
    timeClus1 = sum(clususerDataStat(clus1Ind, 5))/sum(clususerDataStat(:, 5));
    clus2Ind = find(clususerDataStat(:,end) == 2);
    timeClus2 = sum(clususerDataStat(clus2Ind, 5))/sum(clususerDataStat(:, 5));
    clus3Ind = find(clususerDataStat(:,end) == 3);
    timeClus3 = sum(clususerDataStat(clus3Ind, 5))/sum(clususerDataStat(:, 5));
    % Add user id with timeClusters
    gpsFeatures(z,2:4) = [timeClus1, timeClus2, timeClus3];
    % Calculate Entropy
    ent = 0;
    for i = 1:3
        ent = ent + (gpsFeatures(z,i+1) * (log(gpsFeatures(z,i+1))));
    end
    ent = -ent;
    gpsFeatures(z,5) = ent;
    %%
    % Normalized entropy
    gpsFeatures(z,6) = gpsFeatures(z,5)/(log(3));  % N = 3, the total number of clusters
    %%
    % Home stay - cluster which have the time between 12am to 6am
    gpsTime = datestr(clususerDataStat(:,1)/86400 + datenum(1970,1,1));
    t = datetime(gpsTime, 'Format', 'dd/M/yy HH:mm:ss');
    t = datevec(t); % 4th column have hour
    col = size(clususerDataStat,2);
    for i = 1:size(t,1) % Lets home = 0
        if(t(i,4) >= 00 & t(i,4) < 6)
            clususerDataStat(i, col+1) = 0;
        else
            clususerDataStat(i, col+1) = 1;  %Not at home
        end
    end
    % ----------------------------------------------------------------------
    % As of now, our cluster 1 is major but there is an overlap i.e. both
    % cluster 2 and 3 have the home stay time as well - probably because of
    % small campus
    % ----------------------------------------------------------------------
    % To find the percentage of time = total time@Home/total tile
    homeTime = find(clususerDataStat(:,end) == 0);
    gpsFeatures(z,7) = sum(clususerDataStat(homeTime,5))/sum(clususerDataStat(:,5));
    %%

    % Circadian movement ?
    %%
    % Transition time = number of samples in transition/ total number of
    % samples
    movingSamples = size(find(userData(:,4) == 1),1);
    AllSamples = size(find(userData(:,4) ~= 1),1);
    gpsFeatures(z,8) = movingSamples/(movingSamples + AllSamples);
    %%
    % Total distance traveled in km
    sumdistkm = 0;
    for i = 1:(size(userData,1)-1)
        latLong1 = [userData(i,2) userData(i,3)];
        latLong2 = [userData(i+1,2) userData(i+1,3)];
         [d1km d2km]=lldistkm(latLong1,latLong2);
         sumdistkm = sumdistkm + d1km;
    end
    gpsFeatures(z,9) = sumdistkm;
    clear userData timeSpent latDiff longDiff ind userDataStat varLat varLong locVar LongLatStat idx C clususerDataStat clus1Ind clus2Ind clus3Ind timeClus1 timeClus2 timeClus3 i gpsTime homeTime t sumdistkm movingSamples col ent latLong1 latLong2 d1km d2km AllSamples
end
gpsFeatures(:,10) = gpsData{z+1};
%%
% Write down the GPS Features
save('\data\gps_Processesed_feature.mat','gpsFeatures');