OpenMLDB + XGBoost: TalkingData Ad Tracking Fraud Detection#

We will demonstrate how to use OpenMLDB together with other open source software to develop a machine learning application in order to complete the TalkingData AD Fraud Detection challenge (see more information on this challenge in Kaggle).

1 Preparations#

1.1 Download OpenMLDB#

1.1.1 Run in Docker#

It is recommended to run this demo in Docker. Please make sure that OpenMLDB and its dependencies are installed.

Start the OpenMLDB Docker Image

docker run -it 4pdosc/openmldb:0.7.3 bash

1.1.2 Run Locally#

If you want to run this demo locally, please run following command to install all dependencies after downloading OpenMLDB Server pkg and make sure your version is higher than 0.5.0:

pip install pandas xgboost==1.4.2 sklearn tornado "openmldb>=0.5.0" requests

1.2 Data Preparation#

In this example, only the first 10,000 rows of the training data set provided by Kaggle will be used, see the data in train_sample.csv. To download the complete dataset, please run the command:

kaggle competitions download -c talkingdata-adtracking-fraud-detection

After download, please extract the data to demo/talkingdata-adtracking-fraud-detection/data and run the cut_data() function in the train_and_serve.py to make new dataset for training.

1.3 Start the OpenMLDB Cluster#

/work/init.sh

1.4 Start the Prediction Server#

If you have not deployed the prediction server, you can start it with the option --no-init.

python3 /work/talkingdata/predict_server.py --no-init > predict.log 2>&1 &

Tip

  • To stop the prediction server, please use pkill -9 python3 command.

  • You can send the post request <ip>:<port>/update to update the prediction server after training.

2 Training and Serving#

cd /work/talkingdata
python3 train_and_serve.py

The python script train_and_serve.py uses the OpenMLDB for feature extraction and the XGBoost model for training. The script completes the following tasks:

  1. Loading the data into the offline database.

  2. Extracting the features offline:

    • Number of clicks of the ‘ip-day-hour’ in 1h window.

    • Number of clicks of the ‘ip-app’ in unlimited window.

    • Number of clicks of the ‘ip-app-os’ in unlimited window.

  3. Training and saving the model.

  4. Deploying the SQL.

  5. Loading the data into the online storage.

  6. Updating the model on the prediction server.

3 Prediction#

For prediction, send the post request <ip>:<port>/predict to the prediction server, or you can run the Python script through the command below:

python3 predict.py

4 Note#

The pre-installed XGBoost Python wheel may be incompatible with the OpenMLDB Python SDK on your computer, which may lead to the following error: train\_and\_serve.py core dump at SetGPUAttribute...

Installing the XGBoost by the source code may resolve the problem. Please switch to the path of the XGBoost source code and execute: cd python-package && python setup.py install

There is another solution to construct the wheel by executing: python setup.py bdist_wheel