DeepAttribution

DeepAttribution is an AWS (sagemaker) ML pipeline that allows marketing data scientists and ML engineers to compute multi touch-attribution results using the state of the art technique without effort.

Prerequistes

• A big dataset (> 1GB) at impression level (impressions dataset) as parquet file with the following schema:

  • uid: the user/client unique identifier
  • timestamp: unix timestamp
  • campaign: the campaign name
  • conversion: whether or not a conversion happened after the impression

                    uid | timestamp | campaign | conversion 
                    _______________________________________

                    int |    int    |    str   |    bool 

• S3 bucket (deep-attribution bucket) with the following folder hierarchy

    .
    ├── raw                         # contains impressions dataset
    ├── feature_store               # empty before pipeline execution
    ├── feature_store_preprocessed  # empty before pipeline execution
    ├── model                       # empty before pipeline execution
    └── attention_report            # empty before pipeline execution

• Sagemaker notebook instance with git repo set to this repo (deep-attribution bucket).
• AWS role with sagemaker execution permission and read/write permissions on the s3 bucket mentionned above.
• Basic knowledge of LSTM and familiarity with the multi touch attribution model presented in Deep Neural Net with Attention for Multi-channel Multi-touch Attribution
• Understanding of what is a journey, the model and the pipeline

Using DeepAttribution

1. Define the journey maximum length. Please refer to this doc to define it.
2. Update the config file (config.yaml) with the desired instance type and count, bucket name and the journey maximum length.
3. In the deep-attribution instance open the pipeline execution notebook (deep_attribution/pipeline_exec.ipynb)
4. Run all the cells
5. Get the attribution results in the deep-attribution bucket (attention_report/campaign_attention.parquet)

Contact

If you want to contact me you can reach me at leopoldavezac@gmail.com.

License

This project uses the following license: MIT.