AI506 Term Project: E-commerce Product Return Prediction

KAIST 2024 Spring AI506: Data Mining and Search

ABCDE-NET: Attention Block with Customer Dependency for E-commerce NETwork

Project Summary

The ultimate goal of this project is to practice data mining research by predicting product returns in e-commerce settings.

Our method is designed to solve 2 E-commerce Product Return Prediction tasks: (1) Order Return Prediction and (2) Product Return Prediction.

It models complex relations between orders and products as hypergraphs based on the WHATsNET paper: Classification of Edge-Dependent Labels of Nodes in Hypergraphs

The contributions of this project are the following:

• Our proposed method is an effective method for prediction of product returns based on hypergraphs.
• It is effective and comparable to naive methods such as random guessing, SVM, or random forests.
• Initialization of node embeddings from product features (color, size, product group) which are concatenated to random walk initialization.
• Implementation of Customer-dependent MAB for consideration of customer preference on product return.

Team Information

• Dongwon Choi (School of Computing, KAIST)
• Sangwoo Kim (Graduate School of AI, KAIST)
• Dohun Lee (Graduate School of AI, KAIST)

Environment

• Python >= 3.8
• Pytorch: 2.2.1
• Cuda: 11.8
• GPU: RTX 4090

How to run

Train

You can train our model for each task with the following commands:

// Task 1
python train.py --dataset_name task1 --output_dim 3 --vorder_input "degree_nodecentrality,eigenvec_nodecentrality,pagerank_nodecentrality,kcore_nodecentrality" --embedder whatsnet --att_type_v OrderPE --agg_type_v PrevQ --att_type_e OrderPE --agg_type_e PrevQ --num_att_layer 2 --num_layers 1 --scorer sm --scorer_num_layers 1 --bs 1024 --lr 0.001 --sampling 40 --dropout 0.7 --optimizer "adam" --k 0 --gamma 0.99 --dim_hidden 64 --dim_edge 128 --dim_vertex 128 --epochs 100 --test_epoch 1 --evaltype test --save_epochs 1 --seed 42 --fix_seed --recalculate --use_wandb=1 --custom

// Task 2
python train.py --dataset_name task2 --output_dim 2 --vorder_input "degree_nodecentrality,eigenvec_nodecentrality,pagerank_nodecentrality,kcore_nodecentrality" --embedder whatsnet --att_type_v OrderPE --agg_type_v PrevQ --att_type_e OrderPE --agg_type_e PrevQ --num_att_layer 2 --num_layers 1 --scorer sm --scorer_num_layers 1 --bs 1024 --lr 0.001 --sampling 40 --dropout 0.7 --optimizer "adam" --k 0 --gamma 0.99 --dim_hidden 64 --dim_edge 128 --dim_vertex 128 --epochs 100 --test_epoch 1 --evaltype test --save_epochs 1 --seed 42 --fix_seed --recalculate --use_wandb=1 --custom

Predict

You can produce predictions from our model for each task with the following commands:

// Task 1
python predict.py --dataset_name task1 --output_dim 3 --vorder_input "degree_nodecentrality,eigenvec_nodecentrality,pagerank_nodecentrality,kcore_nodecentrality" --embedder whatsnet --att_type_v OrderPE --agg_type_v PrevQ --att_type_e OrderPE --agg_type_e PrevQ --num_att_layer 2 --num_layers 1 --scorer sm --scorer_num_layers 1 --bs 1024 --lr 0.001 --sampling 40 --dropout 0.7 --optimizer "adam" --k 0 --gamma 0.99 --dim_hidden 64 --dim_edge 128 --dim_vertex 128 --epochs 100 --test_epoch 1 --evaltype test --save_epochs 1 --seed 42 --fix_seed --use_wandb=0 --custom

// Task 2
python predict.py --dataset_name task2 --output_dim 2 --vorder_input "degree_nodecentrality,eigenvec_nodecentrality,pagerank_nodecentrality,kcore_nodecentrality" --embedder whatsnet --att_type_v OrderPE --agg_type_v PrevQ --att_type_e OrderPE --agg_type_e PrevQ --num_att_layer 2 --num_layers 1 --scorer sm --scorer_num_layers 1 --bs 1024 --lr 0.001 --sampling 40 --dropout 0.7 --optimizer "adam" --k 0 --gamma 0.99 --dim_hidden 64 --dim_edge 128 --dim_vertex 128 --epochs 100 --test_epoch 1 --evaltype test --save_epochs 1 --seed 42 --fix_seed --use_wandb=0 --custom