MultiBoost benchmark script

This python script shows how to train a multiboost classifier on training.csv, use the score function to order the events, and determine a score threshold on a validation set. The script then classifies the events in test.csv, and creates a file called submission.csv that you can submit at the Kaggle submission site. You can execute this iPython notebook, or you can read the same notebook below (or here in a nicer format).

 

Running the multiboost benchmark on the Higgs boson machine learning challenge

In []:
import random,string,math,csv,pandas
import numpy as np
import matplotlib.pyplot as plt

Train/validation cut and converting into arff

In [19]:
def DataToArff(xs,labels,weights,header,title,fileName):
    outFile = open(fileName + ".arff","w")
    outFile.write("@RELATION " + title + "\n\n")
    #header
    for feature in header:
        outFile.write("@ATTRIBUTE " + feature + 
                      " NUMERIC\n")
    # multiboost requires ±1*weight in two columns
    outFile.write("@ATTRIBUTE classSignal NUMERIC\n")
    outFile.write("@ATTRIBUTE classBackground NUMERIC\n")
    outFile.write("\n@DATA\n")
    for x,label,weight in zip(xs,labels,weights):
        for xj in x:
            outFile.write(str(xj) + ",")
        if label == 's':
            outFile.write(str(weight) + "," + str(-weight) + "\n")
        else:
            outFile.write(str(-weight) + "," + str(weight) + "\n")
    outFile.close()
In [5]:
all = list(csv.reader(open("training.csv","rb"), delimiter=','))
In [9]:
header = np.array(all[0][1:-2])
In [23]:
xs = np.array([map(float, row[1:-2]) for row in all[1:]])
(numPoints,numFeatures) = xs.shape
In [7]:
sSelector = np.array([row[-1] == 's' for row in all[1:]])
bSelector = np.array([row[-1] == 'b' for row in all[1:]])
In [8]:
weights = np.array([float(row[-2]) for row in all[1:]])
labels = np.array([row[-1] for row in all[1:]])
sumWeights = np.sum(weights)
sumSWeights = np.sum(weights[sSelector])
sumBWeights = np.sum(weights[bSelector])

Save/load permutation if you want to use the same cut

In [26]:
randomPermutation = random.sample(range(len(xs)), len(xs))
np.savetxt("randomPermutation.csv",randomPermutation,fmt='%d',delimiter=',')
#randomPermutation = np.array(map(int,np.array(list(csv.reader(open("randomPermutation.csv","rb"), delimiter=','))).flatten()))

90/10 training/validation cut

In [27]:
numPointsTrain = int(numPoints*0.9)
numPointsValidation = numPoints - numPointsTrain

xsTrain = xs[randomPermutation[:numPointsTrain]]
xsValidation = xs[randomPermutation[numPointsTrain:]]

sSelectorTrain = sSelector[randomPermutation[:numPointsTrain]]
bSelectorTrain = bSelector[randomPermutation[:numPointsTrain]]
sSelectorValidation = sSelector[randomPermutation[numPointsTrain:]]
bSelectorValidation = bSelector[randomPermutation[numPointsTrain:]]

weightsTrain = weights[randomPermutation[:numPointsTrain]]
weightsValidation = weights[randomPermutation[numPointsTrain:]]

labelsTrain = labels[randomPermutation[:numPointsTrain]]
labelsValidation = labels[randomPermutation[numPointsTrain:]]

sumWeightsTrain = np.sum(weightsTrain)
sumSWeightsTrain = np.sum(weightsTrain[sSelectorTrain])
sumBWeightsTrain = np.sum(weightsTrain[bSelectorTrain])
In [28]:
DataToArff(xsTrain,labelsTrain,weightsTrain,header,"HiggsML_challenge_train","training")
DataToArff(xsValidation,labelsValidation,weightsValidation,header,"HiggsML_challenge_validation","validation")

Running multiboost training

Download multiboost from here then compile it (or use the precompiled images). Then run
multiboost --configfile config.txt

using config.txt. It trains AdaBoost with 3000 trees of three leaves (two inner nodes). You can change these hyperparameters in the config file:

baselearnertype SingleStumpLearner numInnerNodes
traintest training.arff validation.arff numIterations

Running till the full 3000 iterations will take hours, but you can continue executing this script after about a minute when the code starts to produce the model file shyp.xml and the table with the errors results.dta.

You can plot the learning curves (balanced weighted error rate) any time with the following command.

In [115]:
resultsText = list(csv.reader(open("results.dta","rb"), delimiter='\t'))
ts = [int(result[0]) for result in resultsText]
trainErrors = np.array([float(result[5]) for result in resultsText])
testErrors = np.array([float(result[11]) for result in resultsText])

fig = plt.figure()
fig.suptitle('MultiBoost learning curves', fontsize=14, fontweight='bold')
ax = fig.add_subplot(111)
fig.subplots_adjust(top=0.85)

ax.set_xlabel('number of boosting iterations')
ax.set_ylabel('balanced weighted error rate')

ax.annotate('training error', xy=(0.9*len(ts), trainErrors[len(ts)-1]), 
            xytext=(0.6*len(ts), trainErrors[len(ts)-1]-0.05),
            arrowprops=dict(facecolor='blue', shrink=0.05))
ax.annotate('validation error', xy=(0.9*len(ts), testErrors[len(ts)-1]), 
            xytext=(0.6*len(ts), testErrors[len(ts)-1]+0.05),
            arrowprops=dict(facecolor='red', shrink=0.05))

ax.plot(ts,trainErrors,'b-')
ax.plot(ts,testErrors,'r-')

ax.axis([0, len(ts), 0.1, 0.3])

plt.show()

Optimizing the AMS on the held out validation set

The Approximate Median Significances and b are the sum of signal and background weights, respectively, in the selection region.

In [55]:
def AMS(s,b):
    assert s >= 0
    assert b >= 0
    bReg = 10.
    return math.sqrt(2 * ((s + b + bReg) * 
                          math.log(1 + s / (b + bReg)) - s))

Run

multiboost --configfile configScoresValidation.txt

using configScoresValidation.txt to output the posterior scores scoresValidation.txt on the validation set. You can change the effective number of trees used for the validation score in

posteriors validation.arff shyp.xml scoresValidation.txt numIterations

Loading the scores on the validation set

In [64]:
validationScoresText = list(csv.reader(open("scoresValidation.txt","rb"), delimiter=','))
validationScores = np.array([float(score[0]) for score in validationScoresText])

Sorting the indices in increasing order of the scores.

In [66]:
tIIs = validationScores.argsort()

Weights have to be normalized to the same sum as in the full set.

In [67]:
wFactor = 1.* numPoints / numPointsValidation

Initializing s and b to the full sum of weights, we start by having all points in the selection region. amss will contain AMSs after each point moved out of the selection region in the sorted validation set. amsMax will contain the best validation AMS, and threshold will be the smallest score among the selected points. We will do len(tIIs) iterations, which means that amss[-1] is the AMS when only the point with the highest score is selected.

In [74]:
s = np.sum(weightsValidation[sSelectorValidation])
b = np.sum(weightsValidation[bSelectorValidation])
amss = np.empty([len(tIIs)])
amsMax = 0
threshold = 0.0
for tI in range(len(tIIs)):
    # don't forget to renormalize the weights to the same sum 
    # as in the complete training set
    amss[tI] = AMS(max(0,s * wFactor),max(0,b * wFactor))
    # careful with small regions, they fluctuate a lot
    if tI < 0.9 * len(tIIs) and amss[tI] > amsMax:
        amsMax = amss[tI]
        threshold = validationScores[tIIs[tI]]
        #print tI,threshold
    if sSelectorValidation[tIIs[tI]]:
        s -= weightsValidation[tIIs[tI]]
    else:
        b -= weightsValidation[tIIs[tI]]

Plotting the AMS vs the rank.

In [117]:
fig = plt.figure()
fig.suptitle('MultiBoost AMS curves', fontsize=14, fontweight='bold')
vsRank = fig.add_subplot(111)
fig.subplots_adjust(top=0.85)

vsRank.set_xlabel('rank')
vsRank.set_ylabel('AMS')

vsRank.plot(amss,'b-')

vsRank.axis([0,len(amss), 0, 4])

plt.show()

Plotting the AMS vs the score.

In [118]:
fig = plt.figure()
fig.suptitle('MultiBoost AMS curves', fontsize=14, fontweight='bold')
vsScore = fig.add_subplot(111)
fig.subplots_adjust(top=0.85)

vsScore.set_xlabel('score')
vsScore.set_ylabel('AMS')

vsScore.plot(validationScores[tIIs],amss,'b-')

vsScore.axis([validationScores[tIIs[0]],validationScores[tIIs[-1]] , 0, 4])

plt.show()

Constructing the submission file

Reading the test file, slicing off the header row and the id column, and converting the data into arff.

In [119]:
testText = list(csv.reader(open("test.csv","rb"), delimiter=','))
testIds = np.array([int(row[0]) for row in testText[1:]])
xsTest = np.array([map(float, row[1:]) for row in testText[1:]])
weightsTest = np.repeat(1.0,len(testText)-1)
labelsTest = np.repeat('s',len(testText)-1)
DataToArff(xsTest,labelsTest,weightsTest,header,"HiggsML_challenge_test","test")

Run

multiboost --configfile configScoresTest.txt

using configScoresTest.txt to output the posterior scores scoresTest.txt on the test set. You can change the effective number of tree used for the test score in

posteriors test.arff shyp.xml scoresTest.txt numIterations

Reading the test scores

In [97]:
testScoresText = list(csv.reader(open("scoresTest.txt", "rb"),delimiter=','))
testScores = np.array([float(score[0]) for score in testScoresText])

Computing the rank order.

In [102]:
testInversePermutation = testScores.argsort()
In [103]:
testPermutation = list(testInversePermutation)
for tI,tII in zip(range(len(testInversePermutation)),
                  testInversePermutation):
    testPermutation[tII] = tI

Computing the submission file with columns EventId, RankOrder, and Class.

In [104]:
submission = np.array([[str(testIds[tI]),str(testPermutation[tI]+1),
                       's' if testScores[tI] >= threshold else 'b'] 
            for tI in range(len(testIds))])
In [105]:
submission = np.append([['EventId','RankOrder','Class']],
                        submission, axis=0)

Saving the file that can be submitted to Kaggle.

In [106]:
np.savetxt("submission.csv",submission,fmt='%s',delimiter=',')