Skip to content
An implementation of our CIKM 2018 paper "Deep Conversion Attribution with Dual-attention Recurrent Neural Network"
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.
data update DARNN, add SP May 20, 2018
.gitignore update ignore file. Feb 9, 2018 Update May 8, 2019

Deep Conversion Attribution for Online Advertising

A tensorflow implementation of all the compared models for the CIKM 2018 paper: Learning Multi-touch Conversion Attribution with Dual-attention Mechanisms for Online Advertising.

Paper Link:


Data Preparation

We have uploaded a tiny data sample for training and evaluation in this repository.

The full dataset for this project has been published here.

After downloading please replace the sample data in data/ folder with the full data files.

Data description

Our raw data is Criteo Attribution Modeling for Bidding Dataset . You can download it and read its description on this page.

Below are the descriptions of our data preprocessing.

  1. We group all the impressions by user_id+conversion_id ( regard as one sequence ), shuffle the whole dataset, and then divide it into trainset and testset ( ratio: train 0.8, test 0.2) with negative down sampling (ratio 0.7) at the meanwhile.

  2. We create mapping from features from certain fields ([campaign, cat1, cat2, …, cat9]) to index.

  3. We turn every line into such format: “time click campaign cat1 cat2 … cat9”

Installation and Running

TensorFlow(>=1.2) and dependant packages (e.g., numpy and sklearn) should be pre-installed before running the code.

After package installation, you can simple run the code with the demo tiny dataset.

python [learning rate]                    # for LR
python                                    # for Simple Probablistic
python                                    # for AdditiveHazard
python [learning rate] [batchsize]      # for AMTA
python [learning rate] [batchsize] [mu] # for ARNN
python [learning rate] [batchsize] [mu]# for DARNN

We have set default hyperparameters in the model implementation. So the parameter arguments are optional for running the code.


  title={Learning Multi-touch Conversion Attribution with Dual-attention Mechanisms for Online Advertising},
  author={Ren, Kan and Fang, Yuchen and Zhang, Weinan and Liu, Shuhao and Li, Jiajun and Zhang, Ya and Yu, Yong and Wang, Jun},
  booktitle={Proceedings of the 27th ACM International Conference on Information and Knowledge Management},
You can’t perform that action at this time.