WX challenge

1. 环境依赖

• Python 3.6.5
• numba 0.53.1
• numpy 1.18.5
• pandas 1.0.5
• scikit-learn 0.23.1
• tensorflow-gpu 1.13.1
• tqdm 4.46.1
• scipy 1.5.0
• deepctr 0.8.6
• gensim 3.8

2. 目录结构

./
├── README.md
├── requirements.txt, python package requirements 
├── init.sh, script for installing package requirements
├── train.sh, script for preparing train/inference data and training models, including pretrained models
├── inference.sh, main function for inference on test dataset
├── src
│   ├── prepare, codes for preparing train/inference dataset
|       ├──get_features.py   
│   ├── model, codes for model architecture
|       ├──mmoe.py  
|   ├── train, codes for training 
|       ├──run_submit.py
|   ├── evaluation.py, main function for evaluation 
|   ├── inference.py
|   ├── inference1.py
├── data
│   ├── wedata
|       ├──wechat_algo_data1, dataset of the competition
|       ├──wechat_algo_data2, dataset of the competition
|   ├── submission, prediction result after running inference.sh
|   ├── model, model files
|   ├── feature, feature files

3. 运行流程

• 进入目录：cd /home/tione/notebook/wbdc2021-semi
• 安装环境：使用 conda_tensorflow_py3虚拟环境 运行sh init.sh
• 数据准备和模型训练：sh train.sh
• 预测并生成结果文件：sh inference.sh /home/tione/notebook/wbdc2021-semi/data/wedata/wechat_algo_data2/test_b.csv

4. 模型及特征

• 模型：MMOE
• 参数：
  • batch_size: 4092
  • emded_dim: 512
  • num_epochs: 5
  • learning_rate: 0.01
• 特征：
  • userid, feedid, authorid, bgm_singer_id, bgm_song_id等id类特征
  • keyword、tag标签特征
  • 视频类别、作者类别
  • userid序列embedding
  • feed聚类、author聚类、user聚类

5. 算法性能

• 资源配置：2*P40_48G显存_14核CPU_112G内存
• 预测耗时
  • 总预测时长: 1791 s
  • 单个目标行为2000条样本的平均预测时长: 120.344 ms

6. 代码说明

模型预测部分代码位置如下：

路径	行数	内容
src/inference.py	82 - 96	pred_ans = train_model.predict(test_model_input, batch_size=batch_size * 100)
src/inference1.py	93 - 108	pred_ans = train_model.predict(test_model_input, batch_size=batch_size * 100)

7. 相关文献

• Ma J, Zhao Z, Yi X, et al. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts[C]//Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2018: 1930-1939.
• Weichen Shen. (2017). DeepCTR: Easy-to-use,Modular and Extendible package of deep-learning based CTR models. https://github.com/shenweichen/deepctr.