uic-paper / mimn

Practice on Long Sequential User Behavior Modeling for Click-Through Rate Prediction

Implementation of Practice on Long Sequential User Behavior Modeling for Click-Through Rate Prediction using tensorflow

Prerequisites

• Python 2.x
• Tensorflow 1.4

Data

• Amazon Book Data
  
• Taobao Data

Getting Started

First we need to prepare data.

Amazon Prepare

• You can get the raw Amazon data prepared
  

sh prepare_amazon.sh

• Because getting and processing the data is time consuming，we had processed Amazon data and upload it for you.
  

sh prepare_ready_data.sh

Taobao Prepare

First download Taobao Data to get "UserBehavior.csv.zip", then execute the following command.

sh prepare_taobao.sh

Running

usage: train_book.py|train_taobao.py  [-h] [-p TRAIN|TEST] [--random_seed RANDOM_SEED]
                     [--model_type MODEL_TYPE] [--memory_size MEMORY_SIZE]
                     [--mem_induction MEM_INDUCTION]
                     [--util_reg UTIL_REG]

Base Model

The example for DNN

python script/train_book.py -p train --random_seed 19 --model_type DNN
python script/train_book.py -p test --random_seed 19 --model_type DNN

The model below had been supported:

• DNN
• PNN
• DIN
• GRU4REC
• ARNN
• RUM
• DIEN
• DIEN_with_neg

MIMN

You can train MIMN with different parameter setting:

• MIMN Basic

python script/train_taobao.py -p train --random_seed 19 --model_type MIMN --memory_size 4 --mem_induction 0 --util_reg 0

• MIMN with Memory Utilization Regularization

python script/train_taobao.py -p train --random_seed 19 --model_type MIMN --memory_size 4 --mem_induction 0 --util_reg 1

• MIMN with Memory Utilization Regularization and Memory Induction Unit

python script/train_taobao.py -p train --random_seed 19 --model_type MIMN --memory_size 4 --mem_induction 1 --util_reg 1

• MIMN with Auxiliary Loss

python script/train_taobao.py -p train --random_seed 19 --model_type MIMN_with_neg --memory_size 4 --mem_induction 0 --util_reg 0

If you want to train Amazon Data, you just need replace above train_taobao.py to train_book.py

mimn's Issues

split_by_user.py的脚本有问题吧？

import random

fi = open("local_test", "r")
ftrain = open("local_train_splitByUser", "w")
ftest = open("local_test_splitByUser", "w")

while True:
    rand_int = random.randint(1, 10)
    noclk_line = fi.readline().strip()
    clk_line = fi.readline().strip()
    if noclk_line == "" or clk_line == "":
        break
    if rand_int == 2:
        print >> ftest, noclk_line
        print >> ftest, clk_line
    else:
        print >> ftrain, noclk_line
        print >> ftrain, clk_line

这个脚本对测试集划分为train和test，写的有问题吧？ 不过看起来之前的步骤local_aggretor.py里就已经划分好了吧，

关于训练数据其中的行为类型

请问，我看了在taobao的训练数据中，数据得预处理，其中行为类型好像没有用到？是这样吗？另外时间数据，只是作为一个排序来使用的，并没有考虑行为之间的时间差是吗？

训练集和测试集文档可以放一下嘛？

淘宝数据预处理会卡死？有木有TF2的版本？

UserBehavior.csv共3G，跑了一晚上还是没完成，早上看卡在了3一动不动：
D:\Anaconda3\envs\TF2\python.exe F:/python/MIMN-master/preprocess/taobao_prepare.py
4162024 987994 9439 4
feature_size 4162024 5159462
group completed
987994
get user last touch time completed
1
2
3

然后一个白天也一动不动。有没有童鞋们把mimn跑起来呢？本想改成TF2的版本的，发现_Linear比较麻烦，然后放弃了

"i==1" in mimn.py Line84 should be changed by "i>=self.read_head_num"

I think "i==1" in mimn.py Line84 should be changed by "i>=self.read_head_num".
From my point of view, this "i==1" is used to limit the operation only on write head weight when the self.read_head_num==1 and self.write_head_num==1.
But these two parameter can be changes in the code, so this "i==1" can cause problem when self.read_head_num!=1 or self.write_head_num!=1.