Added functionality to identify extreme events and to make graph of spikeyness vs. error.

Multiplicative Algorithm/Visualizations.py

@@ -18,11 +18,27 @@ def find_mean_trend_particular_weekday(w, weekday):
     return np.mean(w2, axis=0)
 
 
-def plot_particular_weekday(w, weekday, mean = False):
+def find_extreme_events(w, weekday):
+    mean_trend = find_mean_trend_particular_weekday(w, weekday)
+    w = w.reshape(365, 24)
+    w = w[weekday::7]
+    
+    deviation = np.mean(abs(w - mean_trend), axis = 1)/np.mean(w)
+    extreme_weeks = []
+    for week in range(len(deviation)):
+        if deviation[week] > 0.6:
+            extreme_weeks.append(week)
+    return extreme_weeks
+
+def plot_particular_weekday(w, weekday, mean = False, extreme = False):
     for day in range(weekday, 365, 7):
         plot_particular_day(w, day)
     if mean:
         plt.plot(find_mean_trend_particular_weekday(w, weekday), 'ro')
+    if extreme:
+        extreme_weeks = find_extreme_events(w, weekday)
+        for week in extreme_weeks:
+            plt.plot(w[24*7*week + 24*weekday: 24*7*week + 24*weekday +24], 'k-')
 
         
 def plot_all_weekdays(w, mean = False, signature = '??'):
@@ -35,16 +51,16 @@ def plot_all_weekdays(w, mean = False, signature = '??'):
         plt.clf()
 
         
-def plot_all_signatures_particular_weekday(W, weekday, mean = False):
+def plot_all_signatures_particular_weekday(W, weekday, mean = False, extreme = False):
     plt.ioff()
     fig = plt.figure(figsize = (16, 9))  # size in inches        
     for signature in range(config.RANK):
-        plot_particular_weekday(W[:, signature], weekday, mean = mean)
+        plot_particular_weekday(W[:, signature], weekday, mean = mean, extreme = extreme)
         imagename = './Images/Signature_Trends/' + str(signature).zfill(2) + days_of_week[weekday][0:3] + '.png'
         plt.savefig(imagename, dpi=300, bbox_inches='tight')  # dpi ideal for viewing on laptop fullscreen
         plt.clf()
         
-
+        
 def Heatmap(W1, W2):
     assert W1.shape == W2.shape
     Ro = np.corrcoef(W1.T,W2.T)[len(W1.T):,:len(W2.T)]
@@ -215,9 +231,24 @@ def permute_and_sort(W1, H1, W2, permute = False):
     return w1, h1, W2, permutation
     
     
+def spikeyness_vs_error():
+    D = np.loadtxt('D_trips.txt')
+    W = np.loadtxt('W_trips.txt')
+    H = np.loadtxt('H_trips.txt')
+    def spikeyness(link):
+        trend = D[:52*24*7, link].reshape(52,24*7)
+        return np.mean(np.nanstd(trend, axis=0))/np.nanmean(D[:,link])
     
+    def errors():
+        return np.linalg.norm(np.nan_to_num(D - W@H), axis=0)/np.linalg.norm(np.nan_to_num(D), axis=0)
+    
+    plt.figure(figsize=(10,10))
+    plt.plot([spikeyness(link) for link in range(2302)], errors(), '+')
+    plt.xlabel('Deviation from weekly periodicity')
+    plt.ylabel('Error per link')
+    plt.savefig('Spikeyness_vs_Error')
+    return None
     
     
     
     
-    
\ No newline at end of file