# Demo entry 6328100

**Analytics**

Submitted by **anonymous**
on Nov 26, 2016 at 04:41

Language: Python. Code size: 11.9 kB.

# coding: utf-8 # # Why should you care about overfitting? # # #### Learning Objective: In this tutorial, we will learn what it means to overfit a model and how to avoid doing so. # # #### Data: Imagine you own a website and you are able to track how many hits the website receives over time. Your server can only handle a certain amount of requests at a time. # # #### Prediction Question: When you will hit 100,000 hourly hits? # In[1]: #load the libraries and data import numpy as np import scipy as sp data = sp.genfromtxt("C:/Users/tehskhu/Documents/Personal/Analytics/Projects/Building Machine Learning Systems/Data/ch01/data/web_traffic.tsv", delimiter="\t") # Printing the first 10 rows of the data shows us two columns. # 1) the hour - this will be the x variable. # 2) the number of website requests for that hour - this will be the y variable # In[2]: #split the data into x, y x=data[:,0] y=data[:,1] #remove all NAs from both x,y x = x[~sp.isnan(y)] y = y[~sp.isnan(y)] # Let's plot the data to help us visualize what we are dealing with # In[3]: import matplotlib.pyplot as plt get_ipython().magic('matplotlib notebook') plt.figure(1) # plot the (x,y) points with dots of size 7 plt.scatter(x,y,s=7) plt.title("Web traffic over the last month") plt.xlabel("Time") plt.ylabel("Hits/hour") plt.xticks([w*7*24 for w in range(10)], ['week %i' % w for w in range(10)]) #xlabel to be week # plt.autoscale(tight=True) plt.grid(True, linestyle='-', color='0.75') #draw a slightly opaque, dashed grid plt.show() # We will attempt to fit several polynomials to this set of data. But first let's define a way to measure error i.e. how off are we in our prediction with the polynomial vs. the actuals. # The error function below is just the sum of squared differences. # In[4]: def error(f,x,y): return sp.sum((f(x)-y)**2) # ## Let's fit a first order polynomial # In[5]: fp1, residuals, rank, sv, rcond = sp.polyfit(x,y,1, full=True) # The polyfit() function returns the parameters of the fitted model function, fp1 print("Model parameters: %s" % fp1) print(residuals) # We can see that we fit a polynomial of order 1 with the parameters 2.59 and 989.02. This is a straight line of the form y = 2.59x + 989.02 # # We can now form this line using generated x values and then plot it on the same graph to observe the fit in the scatter plot # In[6]: # use poly1d() to create a polynomial function from the model parameters. fp1 is an array of [2.5 989], then poly1d(fp1) is 2.5x + 989. If we had a larger array, it would be x^2 etc. f1 = sp.poly1d(fp1) print(error(f1,x,y)) #plotting the same graph as above plt.figure(2) # plot the (x,y) points with dots of size 7 plt.scatter(x,y,s=7) plt.title("Web traffic over the last month") plt.xlabel("Time") plt.ylabel("Hits/hour") plt.xticks([w*7*24 for w in range(10)], ['week %i' % w for w in range(10)]) #xlabel to be week # plt.autoscale(tight=True) plt.grid(True, linestyle='-', color='0.75') #draw a slightly opaque, dashed grid #adding a straight line into figure 2 fx = sp.linspace(0,x[-1],1000) #generate x-values for plotting plt.plot(fx, f1(fx), linewidth=3) plt.legend(["d=%i" % f1.order], loc="upper left") # We can see that the fit isn't the best... # ## Let's try fitting a second order polynomial # In[7]: f2p = sp.polyfit(x,y,2) print(f2p) f2 = sp.poly1d(f2p) print(error(f2,x,y)) # The quadratic that we fit is y = 1.05x^2 - 5.26x + 1.9*10^3, with an error of 179,983,507. # Plotting this on the graph results in the following. # In[8]: #plotting the initial scatter plot plt.figure(3) # plot the (x,y) points with dots of size 7 plt.scatter(x,y,s=7) plt.title("Web traffic over the last month") plt.xlabel("Time") plt.ylabel("Hits/hour") plt.xticks([w*7*24 for w in range(10)], ['week %i' % w for w in range(10)]) #xlabel to be week # plt.autoscale(tight=True) plt.grid(True, linestyle='-', color='0.75') #draw a slightly opaque, dashed grid #adding a straight line fx = sp.linspace(0,x[-1],1000) #generate x-values for plotting plt.plot(fx, f1(fx), linewidth=3) plt.legend(["d=%i" % f1.order], loc="upper left") #adding the second order polynomial plt.plot(fx, f2(fx), linewidth=3) plt.legend(["d=%i" % f1.order,"d=%i" % f2.order], loc="upper left") # We can see that the fit is getting better. # ## What if we were to try and fit polynomials of orders 3, 5 and 53? # In[9]: f3p = sp.polyfit(x,y,3) f3 = sp.poly1d(f3p) f10p = sp.polyfit(x,y,10) f10 = sp.poly1d(f10p) #plotting the initial scatter plot plt.figure(4) # plot the (x,y) points with dots of size 7 plt.scatter(x,y,s=7) plt.title("Web traffic over the last month") plt.xlabel("Time") plt.ylabel("Hits/hour") plt.xticks([w*7*24 for w in range(10)], ['week %i' % w for w in range(10)]) #xlabel to be week # plt.autoscale(tight=True) plt.grid(True, linestyle='-', color='0.75') #draw a slightly opaque, dashed grid #adding a straight line fx = sp.linspace(0,x[-1],1000) #generate x-values for plotting plt.plot(fx, f1(fx), linewidth=3) #adding the polynomials of order 2, 3 and 10 plt.plot(fx, f2(fx), linewidth=3) plt.plot(fx, f3(fx), linewidth=3) plt.plot(fx, f10(fx), linewidth=3) plt.legend(["d=%i" % f1.order,"d=%i" % f2.order,"d=%i" % f3.order,"d=%i" % f10.order], loc="upper left") # In[10]: print("Error order 1: ", error(f1,x,y)) print("Error order 1: ",error(f2,x,y)) print("Error order 1: ",error(f3,x,y)) print("Error order 1: ",error(f10,x,y)) # We can see that the fit is getting better visually and mathematically. The calculated errors are decreasing as well. This can only be a good thing right? What happens when we keep increasing our polynomial order. Let's go all the way up to order 99. # In[11]: f99p = sp.polyfit(x,y,99) f99 = sp.poly1d(f99p) #plotting the initial scatter plot plt.figure(5) # plot the (x,y) points with dots of size 7 plt.scatter(x,y,s=7) plt.title("Web traffic over the last month") plt.xlabel("Time") plt.ylabel("Hits/hour") plt.xticks([w*7*24 for w in range(10)], ['week %i' % w for w in range(10)]) #xlabel to be week # plt.autoscale(tight=True) plt.grid(True, linestyle='-', color='0.75') #draw a slightly opaque, dashed grid #adding a straight line fx = sp.linspace(0,x[-1],1000) #generate x-values for plotting plt.plot(fx, f1(fx), linewidth=3) #adding the polynomials of order 2, 3 and 10 plt.plot(fx, f2(fx), linewidth=3) plt.plot(fx, f3(fx), linewidth=3) plt.plot(fx, f10(fx), linewidth=3) plt.plot(fx, f99(fx), linewidth=3) plt.legend(["d=%i" % f1.order,"d=%i" % f2.order,"d=%i" % f3.order,"d=%i" % f10.order,"d=%i" % f99.order], loc="upper left") print("Error order 99: ",error(f99,x,y)) # When we run this, we get a warning "polyfit may be poorly conditioned". Even though we had requested an order of 99 for this polynomial, the max that we got was 53. And even with order 53, can you see some problem with the fit. The error has decreased but does the way the line varies across the weeks, how likely is this pattern to continue in the future? Can we be absolutely certain that there are ups and downs within a week? # # We would need some knowledge about website traffic on our specific site to answer these types of questions. But without that, it is fair to see that the line is 'overfiting' the data and will not be predictive of another set of data points. # So what is the solution to this you ask? # # ## We need to split our data into a training set and a testing set. # # The training set is what we used to fit the polynomial on and the testing set acts as a holdout sample that we will use to test our prediction on. # # Before we go on and splitting our dataset, let's make the problem a little more realistic. Let's assume the best fit for our data is actually a polynomial of order 1 for up to 3.5 weeks. After 3.5 weeks there is an inflection point, and we see traffic rise dramatically. So we will model this through two separate polynomials of order 1. # In[12]: inflection = int(3.5*7*24) #calculate the inflection point in hours xa = x[:inflection] #x before inflection xb = x[inflection:] #x after inflection ya = y[:inflection] #y before inflection yb = y[inflection:] #y after inflection fa = sp.poly1d(sp.polyfit(xa,ya,1)) fb = sp.poly1d(sp.polyfit(xb,yb,1)) fa_error = error(fa, xa, ya) fb_error = error(fb, xb, yb) print("Error inflection=%f" % (fa_error + fb_error)) # the total error is the sum of the two polynomials of order 1 each # In[13]: plt.figure(6) plt.scatter(x,y,s=7) plt.title("Web traffic over the last month") plt.xlabel("Time") plt.ylabel("Hits/hour") plt.xticks([w*7*24 for w in range(20)], ['week %i' % w for w in range(10)]) #xlabel to be week # plt.autoscale(tight=True) plt.grid(True, linestyle='-', color='0.75') #draw a slightly opaque, dashed grid #Fits a simple polynomial line of order 1, split at 3.5 weeks plt.plot(xa, fa(xa), linewidth=3) plt.plot(xb, fb(xb), linewidth=3) plt.legend(["d=%i" % fa.order,"d=%i" % fb.order], loc="upper left") plt.show() # The fit for up to 3.5 weeks looks reasonable. For after 3.5 weeks, # # let's split the data into a training and test set. # In[14]: #For data after 3.5 weeks, we will split into 70% train, 30% test frac = 0.7 split_index = int(frac*len(xb)) shuffled = sp.random.permutation(list(range(len(xb)))) #to shuffle the x points since they are currently in order train = sorted(shuffled[split_index:]) test = sorted(shuffled[:split_index]) # We will now do apply the same techniques as earlier. We will fit polynomials of the same degrees and let's see which one performs the best. # In[15]: fbt1 = sp.poly1d(sp.polyfit(xb[train], yb[train], 1)) fbt2 = sp.poly1d(sp.polyfit(xb[train], yb[train], 2)) fbt3 = sp.poly1d(sp.polyfit(xb[train], yb[train], 3)) fbt10 = sp.poly1d(sp.polyfit(xb[train], yb[train], 10)) fbt100 = sp.poly1d(sp.polyfit(xb[train], yb[train], 100)) plt.figure(7) plt.scatter(x,y,s=7) plt.title("Web traffic over the last month") plt.xlabel("Time") plt.ylabel("Hits/hour") plt.xticks([w*7*24 for w in range(10)], ['week %i' % w for w in range(10)]) #xlabel to be week # plt.autoscale(tight=True) plt.grid(True, linestyle='-', color='0.75') #draw a slightly opaque, dashed grid #Fits all the polynomial lines, split at 3.5 weeks plt.plot(xa, fa(xa), linewidth=3) plt.plot(xb[train], fbt1(xb[train]), linewidth=3) plt.plot(xb[train], fbt2(xb[train]), linewidth=3) plt.plot(xb[train], fbt3(xb[train]), linewidth=3) plt.plot(xb[train], fbt10(xb[train]), linewidth=3) plt.plot(xb[train], fbt100(xb[train]), linewidth=3) plt.legend(["d=%i" % fbt1.order,"d=%i" % fbt2.order,"d=%i" % fbt3.order,"d=%i" % fbt10.order,"d=%i" % fbt100.order], loc="upper left") plt.show() # It is hard to tell visually which fit is better. But this is precisely where the test set comes in. We will calculate the error on the test set. # In[16]: #see errors in test set for all 5 models, model with order 2 polynomial performs the best print("Test errors for only the time after inflection point") for f in [fbt1, fbt2, fbt3, fbt10, fbt100]: print("Error d=%i: %f" % (f.order, error(f, xb[test], yb[test]))) # Error 2 is the lowest out of all five models. So that is the best model. # ## Answering our prediction question... When will we hit 100,000 hourly hits? # In[17]: sp.polyfit(xb[train], yb[train], 2) # The polynomial of order 2 that we selected has the following equation: # y = 6.98 * 10^-2 X^2 -74.85 X + 2.18 * 10^4 # Setting y = 100,000 and solving for x will give us that we will hit 100,000 hits / hour at 9.6 weeks.

This snippet took 0.02 seconds to highlight.

Back to the Entry List or Home.