Search-based User Interest Modeling with Lifelong Sequential Behavior Data for Click-Through Rate Prediction
Implementation of Search-based User Interest Modeling with Lifelong Sequential Behavior Data for Click-Through Rate Prediction using tensorflow
- Python 2.x
- Tensorflow 1.15.0
First we need to prepare data.
- Because getting and processing the data is time consuming，we had processed Amazon data and upload it for you.
tar -xzf data.tar.gz
usage: train_taobao_and_book.py [-h] [-mode MODE] [-seed SEED]
[-use_first_att USE_FIRST_ATT]
[-first_att_top_k FIRST_ATT_TOP_K]
[-use_vec_loss USE_VEC_LOSS]
[-long_seq_split LONG_SEQ_SPLIT]
[-short_seq_split SHORT_SEQ_SPLIT]
[-short_model_type SHORT_MODEL_TYPE]
[-long_model_type LONG_MODEL_TYPE]
[-save_iter SAVE_ITER]
[-test_iter TEST_ITER] [-max_len MAX_LEN]
[-seq_len SEQ_LEN]
[-epoch EPOCH] [-memory_size MEMORY_SIZE]
[-batch_size BATCH_SIZE]
[-search_mode SEARCH_MODE] [-level LEVEL]
[-data_type DATA_TYPE]
[-att_func ATT_FUNC]
The example for DNN
python train.py -mode train \
-data_type book \
-max_len 100 \
-short_model_type DIN \
-short_seq_split '90:100' \
-long_model_type DNN \
-long_seq_split '0:90' \
-seed 2 \
-epoch 2 \
-save_iter 10 \
-test_iter 20 \
-search_mode 'None'
The model type below had been supported:
- DNN
- DIN
- MIMN
You can train SIM with two kinds of search unit:
- hard-search
python train.py -mode train \
-data_type book \
-max_len 100 \
-short_model_type DIN \
-short_seq_split '90:100' \
-long_model_type DIN \
-long_seq_split '0:90' \
-seed 2 \
-epoch 2 \
-save_iter 10 \
-test_iter 20 \
-search_mode 'cate' \
-att_func 'dot'
- soft-search
python train.py -mode train \
-data_type book \
-max_len 100 \
-short_model_type DIN \
-short_seq_split '90:100' \
-long_model_type DIN \
-long_seq_split '0:90' \
-seed 2 \
-epoch 2 \
-data_thread_num 5 \
-save_iter 10 \
-test_iter 20 \
-search_mode 'None' \
-use_first_att True \
-first_att_top_k 50 \
-use_vec_loss True \
-att_func 'dot'
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent