oywtece / gme

Graph Meta Embedding (GME) for cold-start ads in CTR/CVR prediction

Graph Meta Embedding (GME)

Graph Meta Embedding (GME) models for generating initial ID embeddings for cold-start ads in CTR/CVR prediction.

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

• Learning Graph Meta Embeddings for Cold-Start Ads in Click-Through Rate Prediction. In SIGIR, ACM, 2021.

arXiv: https://arxiv.org/abs/2105.08909

ACM DL: https://dl.acm.org/doi/10.1145/3404835.3462879

Bibtex

@inproceedings{ouyang2021learning,
  title={Learning Graph Meta Embeddings for Cold-Start Ads in Click-Through Rate Prediction},
  author={Ouyang, Wentao and Zhang, Xiuwu and Ren, Shukui and Li, Li and Zhang, Kun and Luo, Jinmei and Liu, Zhaojie and Du, Yanlong},
  booktitle={SIGIR},
  pages={1157--1166},
  year={2021},
  organization={ACM}
}

TensorFlow (TF) version

1.12.0

Abbreviation

ft - feature, slot == field

Data Preparation

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

• Each csv row contains the label and then the features (first one-hot features and then multi-hot features).

Assume there are N unique fts. 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 title).

We need to prepare two datasets: one for the main CTR prediction model (including pre-training and warm-up training) and the other for the GME model (i.e., embedding generation model).

1) Dataset for the main CTR prediction model

One row of the csv data looks like:

• <label><one-hot fts><mul-hot fts>

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

• n_one_hot_slot: num of one-hot fts
• n_mul_hot_slot: num of mul-hot fts
• max_len_per_slot: max num of fts in each mul-hot slot

Example 1

1. original fts (ft_name:ft_value)

• label:0, user_id:a, movie_id:b, title:c/d/e

2. csv fts

• 0, a, b, c, d, e, 0, 0

Explanation 1

csv format:
<label><one-hot fts><mul-hot fts>

csv format settings:
n_one_hot_slot = 2 # user_id & movie_id
n_mul_hot_slot = 1 # title
max_len_per_slot = 5

For the mul-hot ft slot "title", we have 3 fts, which are c, d and e. Terefore, we pad 2 zeros (because max_len_per_slot = 5). If there are more than max_len_per_slot fts, we keep only the first max_len_per_slot.

Differences btw pre-training and warm-up training

• Pre-training: use labeled old ads; the aim is to obtain the embedding matrix and FC layer weights
• Warm-up training: use a small number of labeled new ads; the aim is to obtain updated embeddings for new ad IDs (but keep other params unchanged)

2) Dataset for the GME model (use meta learning)

One row of the csv data looks like:

• <label><one-hot fts><mul-hot fts><ngb1 one-hot fts><ngb1 mul-hot fts><ngb2 one-hot fts><ngb2 mul-hot fts> ...

Example 2

1. original fts (ft_name:ft_value)

• label:0, user_id:a, movie_id:b, title:c/d/e, ngb_1_user_id:i, ngb_1_movie_id:j, ngb_1_title:k/l, ngb_2_user_id:s, ngb_2_movie_id:t, ngb_2_title:x/y/z/c/d/e

2. csv fts

• 0, a, b, c, d, e, 0, 0, i, j, k, l, 0, 0, 0, s, t, x, y, z, c, d

Explanation 2

csv format:
<label><one-hot fts><mul-hot fts><ngb1 one-hot fts><ngb1 mul-hot fts><ngb2 one-hot fts><ngb2 mul-hot fts> ...

csv format settings:
n_one_hot_slot = 2 # user_id & movie_id
n_mul_hot_slot = 1 # title
max_len_per_slot = 5

n_one_hot_slot_ngb = 2 # user_id & movie_id
n_mul_hot_slot_ngb = 1 # title
max_len_per_slot_ngb = 5

Sample Data

In the data folder.
Reformatted MovieLens-1M data (csv/tfrecord format with ft index).
The scripts run much faster with the tfrecord data files. We provide tfrecord_writer_new.py which can easily convert csv files to tfrecord files.

• We use 8 features: user_id, movie_id, gender, age, occ, release year, title (multi-hot), genres (multi-hot)

• The following csv files contains 1 + 16 cols (1 label + 6 one-hot slot + 2 multi-hot slots * 5 values each)
  big_train_main.csv
  test_oneshot_a.csv
  test_oneshot_b.csv
  test_oneshot_c.csv
  test_test.csv

• The following csv files contains 1 + 16 + 160 cols, with 10 neighbors appended (acc. to the target movie_id)
  train_oneshot_a_w_ngb.csv
  train_oneshot_b_w_ngb.csv
  train_oneshot_c_w_ngb.csv
  train_oneshot_d_w_ngb.csv
  test_test_w_ngb.csv

Get Sample Data Ready

Go to the data folder.

unzip csv.zip
bash run_tfrecord_writer.sh

Get Pre-trained DNN Model Ready

Go to the tmp folder.

unzip dnn_1011_1705.zip

Config

Validation and hyperparameter tuning

You can set multiple values for hyperparameters.
Example: eta_range = [0.01, 0.02, 0.05, 0.1, 0.2, 0.5]

Testing

Set only 1 value for the optimal hyperparameter found in the validation step.
Example: eta_range = [0.02]

Source Code

• config_dnn.py -- config file for DNN (main CTR prediction model)
• config_gme.py -- config file for GME models (cold-start embedding generation model)
• ctr_funcs.py -- functions
• dnn.py -- DNN model
• gme_all_in_one.py -- GME model, no warm-up training (meta_mode can be set to: 'self', 'ngb', 'GME-P', 'GME-G', 'GME-A')
• gme_all_in_one_warm_up.py - GME model, with warm-up training (need to set warm_up_bool=True)

Run the Code

Train & Test GME

nohup python gme_all_in_one.py > gme_[output_file_name].out 2>&1 &

Train & Test DNN

• This step is not necessary because one saved DNN model is already provided in the 'tmp' folder.

nohup python dnn.py > dnn_[output_file_name].out 2>&1 &