W/Z classification
Estimated reading time: 2 minutes You may find the project at http://cepcgit.ihep.ac.cn/zhuyongfeng/ee_zz_vvqq, use
git clone git@cepcgit.ihep.ac.cn:zhuyongfeng/ee_zz_vvqq.git
to have a local copy of this project.
Introduction
 When z and w all decay to quarks, and these quarks will hadronization. We want to classify w and z through using the information of their decay products. In this experiment, we will use the four momentum of quarks decayed from w or z to classify w and z.
Method
There are two ways to use the information of four momentum of quarks:
 Using the information of invariant mass calculated through correctly pairing two quarks among four quarks decayed from two w or z bosons. We all know the restmass of w and z is 80Gev and 91Gev separately. So the value of invariant mass will help us classify w and z.
 Using deepneuralnetwork to classify w and z.
Samples:

For this exprement, the w and z boson are product from the collision of position and electron, it is in generation level, not through simulation or reconstruction. The channels we use are as following table:
The definition of accuracy

We want to classify w and z by two ways independently, so need to make a rule to compare the classification result of these two ways. In this experiment, we use the following rule: Assumpt the number of w and z boson is wtruth and ztruth separately, through the way of invariant mass or deep learning, correctly classifying the number of w and z boson is wpre and zpre separately, since that there are two classes, so the incorrectly classifying number of w and z boson is (wtruthwpre) and (ztruthzpre) separately. The definition is as following table:
Invariantmass method

Since the charge of z and w boson is zero and one separately. We can use this character to correctly pairing quarks to calculate invariant mass. The following graph is the distribution of invariant mass of w and z:

The following graph is the classification accuracy:
Deeplearning method
 We all know deeplearning is powerful. For out problem, the convolutionneuralnetwork and recurrentneuralnetwork donâ€™t work well since there is no structure information or sequential information, so we decide to use deepneuralnetwork to classify w and z boson.
 The samples are divided into three sets: training set, validation set and testing set. For out experiment, there are 600000 samples in training set, 100000 samples in validation set and 80000 samples in testing set.
 Each sample have several features, the concrete number and type of features is due to our selection. These features will help us to classify w and z.
 The structure of DNN model is shown in http://cepcgit.ihep.ac.cn/zhuyongfeng/ee_zz_vvqq.
Deeplearning result

There are four conditions as show in following table as well as their classification result:
Conclusion
 You have to say that deeplearning is powerful.