Replication of Deep & Cross Network Paper (2017)

This repo replicates the Deep & Cross Network for Ad Click Predictions published by Wang et al. (2017).

Model Architecture

Architecture of the complete DCN

source: Figure 1 Wang et al. (2017)

Architecture of single cross network level source: Figure 2 Wang et al. (2017)

Data

The paper uses data in the Display Advertising Challenge on kaggle, which uses a Criteo dataset.

The data set is no longer available on kaggle, but is available on huggingface

Download link

Other papers on this data set can be found here

Install

See getting_started.sh.

You will need to install poetry if you don't have it. See https://python-poetry.org/docs/.

Without poetry

If you don't want to use poetry, you can try to manually install the python packages listed under [tool.poetry.dependencies] in pyproject.toml. For GPU usage, do pip install tensorflow[and-cuda] instead of pip install tensorflow.

Using poetry

For GPU enabled.

    `poetry install`

For no GPU enabled.

    `poetry install --without nvidia`

To enable plotting DAGs

To enable plotting of neural network graph you need graphvis.

    sudo apt-get install graphviz

Run

Download the data. See above. Then run preprocessing. This will save vocabulary files in data/vocab, and save some plots you can view here

    `python preprocessing.py`  

Start up tensorboard to monitor the training. save_path is the argument passed to DeepCrossNetwork in main.py

    `tensorboard --logdir=./logs/tensorboard/{save_path}`  

Run the training. Perhaps modify the input file from data/train.csv to data/sample.csv to do an initial test run (make sure to also remove the validation file for the test).

    `python main.py`