本项目的数据集给出了15万左右的用户以及12万左右的商品, 以及对应的经过脱敏处理的用户特征和经过预处理的商品特征，旨在为用户推荐50个其可能购买的商品。
本项目采用传统的召回＋排序的方案。在召回模块采用deepwalk， node2vec，item_feature, itemCF四种方法进行多路召回，为每位用户召回1000个商品。在排序阶段采用wide&deep模型，对召回的1000个商品进行排序。将排序所得的分数依据商品点击量进行后处理，来增大对非热门商品的曝光度。最后根据处理后的分数为每位用户推荐50个商品。
最终实现了在验证集上top50召回率0.807， 测试集上top50召回率0.712
数据来源于阿里天池平台开源数据，在百度网盘里面，可以自行下载，按照以下路径创建文件夹以及放置数据。
百度网盘链接：https://pan.baidu.com/s/1sspNWKYVxf-QFTrCjdqfoQ 提取码：853t
│ feature_list.csv # List the features we used in ranking process
│ project_structure.txt # The tree structure of this project
├─ build_graph_model.py # Build deepwalk model and node2vec model
├─ final_rank.py # Build wide&deep network
├─ final_solution.py # Main program
├─ recall_function.py # Functions used to recall items
├─ item_feat.pkl # Item feature after PCA
├─ top100_recall_feature.pkl # Recalled 100 items for each user by using item_feature
├─ top300_recall_deepwalk_result.pkl # Recalled 300 items for each user by using deepwalk
├─ top300_recall_node2vec_result.pkl # Recalled 300 items for each user by using node2vec
├─ topk_recall.pkl # Recalled 1000 items for each user by combining all ways
├─ train_eval_rank.pkl # Cross validation set after ranking
├─ wide_and_deep.h5 # Wide&Deep model using full training set
├─ wide_and_deep_no_cv.h5 # Wide&Deep model using training set except cross validation set
├─ data # Origin dataset
│ ├─ underexpose_test
│ └─ underexpose_train
├─ readme.md
├─ deepwalk_offline.bin # deepwalk model
└─ node2vec_offline.bin # node2vec model
tensorflow==2.3.1
scikit-learn==0.23.2
joblib==0.17.0
networkx==2.1
gensim==3.8.3
pandas==0.25.1
numpy==1.18.5
tqdm==4.26.0
本项目所有代码仅供各位同学学习参考使用。
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent