Mcformer---An end to end model to predict customer-purchase-intention

发明人： Jiahui Feng(冯家辉), Qihang Zhao(赵启航), Hefu Liu(刘和福), Wei Chen(陈伟)

• This algorithm was developed by members of the IOM Lab at the School of Management, USTC
• This is an end to end deep learning method Mcformer to utilize the customer clickstream data to predict the user purchase intention.
• We aim to utilize the customer clickstream data to predict the customer purchase intention, the scenes as follows:
• 

The framework of Mcformer

• Introduction of Mcformer
• In order to deal with multi-dimension clickstream sequence data, we proposed an end-to-end deep learning model, named Multi-channel for purchase transformer (Mcformer), to predict the customers’ purchasing intention. Figure 1 shows the model architecture of Mcformer. This model composed by four parts: embedding layer, multi-transformer layer, cross fusion layer and output layer. Embedding layer is used to embed the sparse one-hot vectors of the behavior data to dense vectors. After that, multi-channels transformer identify intra-information of each sequence. Then the cross fusion layer is applied to identify the inter-information of different sequences. Finally, Mcformer output the result by the multilayer perceptron.

Requirements

• sklearn
• pandas
• pytorch
• cuda

Run the project

''' bash python train.py '''

Data

Our data is the real world data from https://tianchi.aliyun.com/dataset/dataDetail?dataId=649, this dataset need to preprocessing, which need long time. If you need the data to verify our model, you could contact with us jiahuifeng@ustc.mail.edu.cn

• 

modify files to ensure the code work

And if you want to use your data, you have to provide

• the number of items: ni
• the number of category:nc
• the number of types: nt
• the number of hour/minutes: nh

you need to modify the files as follows:

modelconf.py

• cat_pad_unk = [[0,nc+1,nc+2], [0,nh+1,nh+2]]
• item_pad_unk =[[0,ni+1,ni+2], [0,nh+1,nh+2]]
• type_pad_unk =[[0,nt+1,nt+2], [0,nh+1,nh+2]]

data.py

the file location

train.py

if you want to visualize the training process ,you should change parameters.

The results

• The result show that Mcformer get great performance in long sequence classification tasks.
• The beat parameters in the file named optimizpara, so you can use our best parameters directly.

Parameter

• We have disclosed the model best parameters, due to the large file size, you can click the hyperlink to download:
• HyperLink:https://pan.baidu.com/s/10pkew7_tZgbdbjGm0XxDbA verification code：pxxk