rec-pangu 0.3.5

Some Rank/Multi-task model implemented by Pytorch

Rec PanGu

1.开源定位

• 使用pytorch对经典的rank/多任务模型进行实现，并且对外提供统一调用的API接口，极大的降低了使用Rank/多任务模型的时间成本
• 该项目使用了pytorch来实现我们的各种模型，以便于初学推荐系统的人可以更好的理解算法的核心思想
• 由于已经有了很多类似的优秀的开源，我们这里对那些十分通用的模块参考了已有的开源，十分感谢这些开源贡献者的贡献

2.安装

#最新版
git clone https://github.com/HaSai666/rec_pangu.git
cd rec_pangu
pip install -e . --verbose

#稳定版 
pip install rec_pangu --upgrade

3.Rank模型

模型	论文	年份
WDL	Wide & Deep Learning for Recommender Systems	2016
DeepFM	DeepFM: A Factorization-Machine based Neural Network for CTR Prediction	2017
NFM	Neural Factorization Machines for Sparse Predictive Analytics	2017
FiBiNet	FiBiNET: Combining Feature Importance and Bilinear Feature Interaction for Click-Through Rate	2019
AFM	Attentional Factorization Machines	2017
AutoInt	AutoInt: Automatic Feature Interaction Learning via Self-Attentive Neural Networks	2018
CCPM	A Convolutional Click Prediction Model	2015
LR	/	2019
FM	/	2019
xDeepFM	xDeepFM: Combining Explicit and Implicit Feature Interactions for Recommender Systems	2018
DCN	Deep & Cross Network for Ad Click Predictions	2019
MaskNet	MaskNet: Introducing Feature-Wise Multiplication to CTR Ranking Models by Instance-Guided Mask	2021

4.多任务模型

模型	论文	年份
MMOE	Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts	2018
ShareBottom	Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts	2018
ESSM	Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate	2018
OMOE	Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts	2018
MLMMOE	/	/
AITM	Modeling the Sequential Dependence among Audience Multi-step Conversions with Multi-task Learning in Targeted Display Advertising	2019

5.序列召回模型

目前支持如下类型的序列召回模型:

• 经典序列召回模型
• 基于图的序列召回模型
• 多兴趣序列召回模型

模型	类型	论文	年份
YotubeDNN	经典序列召回	Deep Neural Networks for YouTube Recommendations	2016
Gru4Rec	经典序列召回	Session-based recommendations with recurrent neural networks	2015
Narm	经典序列召回	Neural Attentive Session-based Recommendation	2017
NextItNet	经典序列召回	A Simple Convolutional Generative Network for Next Item	2019
ComirecSA	多兴趣召回	Controllable multi-interest framework for recommendation	2020
ComirecDR	多兴趣召回	Controllable multi-interest framework for recommendation	2020
Mind	多兴趣召回	Multi-Interest Network with Dynamic Routing for Recommendation at Tmall	2019
Re4	多兴趣召回	Re4: Learning to Re-contrast, Re-attend, Re-construct for Multi-interest Recommendation	2022
CMI	多兴趣召回	mproving Micro-video Recommendation via Contrastive Multiple Interests	2022
SRGNN	图序列召回	Session-based Recommendation with Graph Neural Networks	2019
GC-SAN	图序列召回	SGraph Contextualized Self-Attention Network for Session-based Recommendation	2019
NISER	图序列召回	NISER: Normalized Item and Session Representations to Handle Popularity Bias	2019

6.图协同过滤模型

TODO

7.Demo

我们的Rank和多任务模型所对外暴露的接口十分相似,同时我们这里也支持使用wandb来实时监测模型训练指标,我们下面会分别给出Rank,多任务模型,wandb的demo

7.1 排序任务Demo

# 声明数据schema
import torch
from rec_pangu.dataset import get_dataloader
from rec_pangu.models.ranking import WDL, DeepFM, NFM, FiBiNet, AFM, AFN, AOANet, AutoInt, CCPM, LR, FM, xDeepFM
from rec_pangu.trainer import RankTrainer
import pandas as pd

if __name__ == '__main__':
    df = pd.read_csv('sample_data/ranking_sample_data.csv')
    print(df.head())
    # 声明数据schema
    schema = {
        "sparse_cols": ['user_id', 'item_id', 'item_type', 'dayofweek', 'is_workday', 'city', 'county',
                        'town', 'village', 'lbs_city', 'lbs_district', 'hardware_platform', 'hardware_ischarging',
                        'os_type', 'network_type', 'position'],
        "dense_cols": ['item_expo_1d', 'item_expo_7d', 'item_expo_14d', 'item_expo_30d', 'item_clk_1d',
                       'item_clk_7d', 'item_clk_14d', 'item_clk_30d', 'use_duration'],
        "label_col": 'click',
    }
    # 准备数据,这里只选择了100条数据,所以没有切分数据集
    train_df = df
    valid_df = df
    test_df = df

    # 声明使用的device
    device = torch.device('cpu')
    # 获取dataloader
    train_loader, valid_loader, test_loader, enc_dict = get_dataloader(train_df, valid_df, test_df, schema)
    # 声明模型,排序模型目前支持：WDL, DeepFM, NFM, FiBiNet, AFM, AFN, AOANet, AutoInt, CCPM, LR, FM, xDeepFM
    model = xDeepFM(enc_dict=enc_dict)
    # 声明Trainer
    trainer = RankTrainer(num_task=1)
    # 训练模型
    trainer.fit(model, train_loader, valid_loader, epoch=5, lr=1e-3, device=device)
    # 保存模型权重
    trainer.save_model(model, './model_ckpt')
    # 模型验证
    test_metric = trainer.evaluate_model(model, test_loader, device=device)
    print('Test metric:{}'.format(test_metric))

7.2 多任务模型Demo

import torch
from rec_pangu.dataset import get_dataloader
from rec_pangu.models.multi_task import AITM, ShareBottom, ESSM, MMOE, OMOE, MLMMOE
from rec_pangu.trainer import RankTrainer
import pandas as pd

if __name__ == '__main__':
    df = pd.read_csv('sample_data/multi_task_sample_data.csv')
    print(df.head())
    #声明数据schema
    schema = {
        "sparse_cols": ['user_id', 'item_id', 'item_type', 'dayofweek', 'is_workday', 'city', 'county',
                        'town', 'village', 'lbs_city', 'lbs_district', 'hardware_platform', 'hardware_ischarging',
                        'os_type', 'network_type', 'position'],
        "dense_cols": ['item_expo_1d', 'item_expo_7d', 'item_expo_14d', 'item_expo_30d', 'item_clk_1d',
                       'item_clk_7d', 'item_clk_14d', 'item_clk_30d', 'use_duration'],
        "label_col": ['click', 'scroll'],
    }
    #准备数据,这里只选择了100条数据,所以没有切分数据集
    train_df = df
    valid_df = df
    test_df = df

    #声明使用的device
    device = torch.device('cpu')
    #获取dataloader
    train_loader, valid_loader, test_loader, enc_dict = get_dataloader(train_df, valid_df, test_df, schema)
    #声明模型,多任务模型目前支持：AITM,ShareBottom,ESSM,MMOE,OMOE,MLMMOE
    model = AITM(enc_dict=enc_dict)
    #声明Trainer
    trainer = RankTrainer(num_task=2)
    #训练模型
    trainer.fit(model, train_loader, valid_loader, epoch=5, lr=1e-3, device=device)
    #保存模型权重
    trainer.save_model(model, './model_ckpt')
    #模型验证
    test_metric = trainer.evaluate_model(model, test_loader, device=device)
    print('Test metric:{}'.format(test_metric))

7.3 序列召回Demo

import torch
from rec_pangu.dataset import get_dataloader
from rec_pangu.models.sequence import ComirecSA,ComirecDR,MIND,CMI,Re4,NARM,YotubeDNN,SRGNN
from rec_pangu.trainer import SequenceTrainer
from rec_pangu.utils import set_device
import pandas as pd

if __name__=='__main__':
    #声明数据schema
    schema = {
        'user_col': 'user_id',
        'item_col': 'item_id',
        'cate_cols': ['genre'],
        'max_length': 20,
        'time_col': 'timestamp',
        'task_type':'sequence'
    }
    # 模型配置
    config = {
        'embedding_dim': 64,
        'lr': 0.001,
        'K': 1,
        'device':-1,
    }
    config['device'] = set_device(config['device'])
    config.update(schema)

    #样例数据
    train_df = pd.read_csv('./sample_data/sample_train.csv')
    valid_df = pd.read_csv('./sample_data/sample_valid.csv')
    test_df = pd.read_csv('./sample_data/sample_test.csv')

    #声明使用的device
    device = torch.device('cpu')
    #获取dataloader
    train_loader, valid_loader, test_loader, enc_dict = get_dataloader(train_df, valid_df, test_df, schema, batch_size=50)
    #声明模型,序列召回模型模型目前支持： ComirecSA,ComirecDR,MIND,CMI,Re4,NARM,YotubeDNN,SRGNN
    model = ComirecSA(enc_dict=enc_dict,config=config)
    #声明Trainer
    trainer = SequenceTrainer(model_ckpt_dir='./model_ckpt')
    #训练模型
    trainer.fit(model, train_loader, valid_loader, epoch=500, lr=1e-3, device=device,log_rounds=10,
                use_earlystoping=True, max_patience=5, monitor_metric='recall@20',)
    #保存模型权重和enc_dict
    trainer.save_all(model, enc_dict, './model_ckpt')
    #模型验证
    test_metric = trainer.evaluate_model(model, test_loader, device=device)

7.4 图协调过滤Demo

TODO