MiNet

Mixed Interest Network (MiNet) is a model for cross-domain click-through rate (CTR) prediction.

If you use this code, please cite the following paper:

MiNet: Mixed Interest Network for Cross-Domain Click-Through Rate Prediction. In CIKM, ACM, 2020.

arXiv: https://arxiv.org/abs/2008.02974

ACM DL: https://dl.acm.org/doi/10.1145/3340531.3412728

Bibtex

@inproceedings{ouyang2020minet,
  title={MiNet: Mixed Interest Network for Cross-Domain Click-Through Rate Prediction},
  author={Ouyang, Wentao and Zhang, Xiuwu and Zhao, Lei and Luo, Jinmei and Zhang, Yu and Zou, Heng and Liu, Zhaojie and Du, Yanlong},
  booktitle={CIKM},
  pages={2669--2676},
  year={2020},
  organization={ACM}
}

TensorFlow (TF) version

1.12.0

Abbreviation

ft - feature, src - source, tar - target, slot == field

Data Preparation

Data is in the "csv" or the "tfrecord" format.

Each csv row contains the label, the user, the target item, clicked items in the target domain and clicked items in the source domain.
In the Amazon example here, the target domain is the Movies&TV and the source domain is the Books.
In UC Toutiao, the target domain is the Ads and the source domain is the News.

Assume there are N unique fts across domains (where user fts are shared). Fts need to be indexed from 1 to N. Use 0 for missing values or for padding.

We categorize fts as i) one-hot or univalent (e.g., user id, city) and ii) mul-hot or multivalent (e.g., words in ad title).

We need to prepare two data sets. One for the target domain, and the other for the source domain.

We need to specify the following parameters for data partitioning after the data are loaded:

n_user_slot [num of user ft slots; it must be the same in 2 data sets]
n_one_hot_slot_1 [num of one-hot fts in the items in the target domain]
n_mul_hot_slot_1 [num of mul-hot fts in the items in the target domain]
max_len_per_slot_1 [max num of fts in each mul-hot slot in the target domain]
max_n_clk_ori_1 [max num of clicked items in the target domain]
n_one_hot_slot_2 [num of one-hot fts in the items in the source domain]
n_mul_hot_slot_2 [num of mul-hot fts in the items in the source domain]
max_len_per_slot_2 [max num of fts in each mul-hot slot in the source domain]
max_n_clk_ori_2 [max num of clicked items in the source domain]

Target domain dataset

Assume we have at most 2 clicked items in the target domain (tar_clk) and 2 clicked items in the source domain (src_clk), then one row of the csv data looks like:

<label><user fts><tar one-hot fts, tar mul-hot fts><tar_clk_1 one-hot fts, tar_clk_1 mul-hot fts><tar_clk_2 one-hot fts, tar_clk_2 mul-hot fts><src_clk_1 one-hot fts, src_clk_1 mul-hot fts><src_clk_2 one-hot fts, src_clk_2 mul-hot fts>

Example 1 (target item)

original fts (ft_name:ft_value)

label:0, user_id:u_1, item_id:i_1, title:apple, title:fruit

csv fts

0, u_1, i_1, apple, fruit, 0

Explanation 1

csv format:
<label><user fts><tar one-hot fts, tar mul-hot fts>

csv format settings:
n_user_slot = 1 # user_id
n_one_hot_slot_1 = 1 # item_id
n_mul_hot_slot_1 = 1 # title
max_len_per_slot_1 = 3

For the mul-hot ft slot "title", we have 2 fts, which are "apple" and "fruit". Terefore, we pad 1 zero (because max_len_per_slot_1 = 3). If there are more than max_len_per_slot_1 fts, we keep only the first max_len_per_slot_1.

Example 2 (target item, target domain clicked item 1, target domain clicked item 2)

original fts (ft_name:ft_value)

label:0, user_id:u_1, item_id:i_1, title:apple, title:fruit; item_id:i_2, title:a, title:b, title:c, title:d; item_id:i_3, title:e

csv fts

0, u_1, i_1, apple, fruit, 0, i_2, a, b, c, i_3, e, 0, 0

Explanation 2

csv format:
<label><user fts><tar one-hot fts, tar mul-hot fts><tar_clk_1 one-hot fts, tar_clk_1 mul-hot fts><tar_clk_2 one-hot fts, tar_clk_2 mul-hot fts>

Source domain dataset

One row of the csv data looks like:

<label><user fts><src one-hot fts, src mul-hot fts>

Sample Data

In the data folder.
Reformatted Amazon data (csv/tfrecord format with ft index).
The scripts run much faster with the tfrecord data files. We provide tfrecord_writer.py which can easily convert csv files to tfrecord files.

Target domain: Movies&TV, Source domain: Books.
Target domain data files: Movies_Books_xx.tfrecord (xx is in {train, val, test})
Source domain data files: Books_xx.tfrecord (xx is in {train, val, test})
Please refer to the paper for more details about the datasets.

Original Amazon datasets:

Original Amazon datasets

Config

Validation and hyperparameter tuning

train file: xx_train.tfrecord, test file: xx_val.tfrecord
Set multiple values for hyperparameters.
Example: eta_range = [0.01, 0.02, 0.05, 0.1, 0.2, 0.5]

Testing

train files: xx_train.tfrecord & xx_val.tfrecord, test file: xx_test.tfrecord
Set only 1 value for the optimal hyperparameter found in the validation step.
Example: eta_range = [0.02]

Source Code

config_amazon.py -- config file
ctr_funcs.py -- functions
dnn_para_tune.py -- DNN model
minet_para_tune.py -- MiNet model

Run the Code

x_para_tune.py files perform hyperparameter tuning. Config the hyperparameter ranges in config_amazon.py.

Run the code

nohup python dnn_para_tune.py > dnn_[output_file_name].out 2>&1 &
nohup python minet_para_tune.py > minet_[output_file_name].out 2>&1 &

forchchch / minet_released_code