Skip to content

Latest commit

 

History

History
50 lines (40 loc) · 3.55 KB

README.md

File metadata and controls

50 lines (40 loc) · 3.55 KB

Kaggle TalkingData AdTracking Fraud Detection Challenge

Predict whether or not an Ad click collected by TalkingData is a click fraud or not based on the relevant information about the click (device, OS, IP, timestamp, and etc). Link to the competition page Kaggle TalkingData. The jupyter notebook talkingData_EDA.ipynb contains preliminary analysis and visualization of the dataset.

Authors: Yunkun Xie (honeyo0obadger@gmail.com), Jianhua Ma (jm9yq@virginia.edu)

Installation

Download the data

  • Clone this repo to your computer (it comes with small sample inputs for test/debug purposes).
  • Download the complete dataset from Kaggle here and save them into the inputs folder (be cautious that the datasets are big).
  • switch into the model folder cd model.

Install the requirements

  • Install the requirements using pip install -r requirements.txt.
  • Make sure to use the Python 3 version (we recommend using Anaconda python distribution).
  • It is recommended to use virtual environment to have a clean package setup.

Usage

We use click package to create simple CLI for the modeling tasks. Each task is created as an indepndent command and all tasks are grouped in the main.py file for simplicity.

  • Run mkdir data_dir to create a directory for processed and engineered data.
  • Run python main.py to see what commands (task) are available.
  • Run python main.py COMMAND [ARGS] to run each task.

Example

Here we show a simple workflow as an example of how to run the script. Assume the raw data files are put under the */inputs directory (here we use sample inputs for fast calculations) and the python scripts are under */model folder. Two separate folders are also created */data and */outputs for processed/engineered dataset and predictions respectively. Notice all the folders are on the same level.

  • Run cd model to switch into model folder.
  • Run python main.py preprocess ../inputs/train_sample.csv ../inputs/test_sample.csv ../data/ to preprocess the raw dataset and save processed dataset in ../data/ folder.
  • Run python main.py feature_engineer ../data/ to generate engineered features and save under the same ../data/ folder. Due to the large size and long processing time of each feature, they are saved in individual files so that they can be assembled later for testing different combinations of features without regenerating them.
  • Run python main.py model_train_predict ../data/ ../outputs/ to train the model and make predictions on the test set. The predictions are saved under ../outputs/ folder.

Extending this

Building a good model requires a lot of testing and adjusting. Current scripts offer the basic (but complete) workflow for the modeling process. It can be easily extended due to its modular nature. For example, each engineered feature is created as a separate function put in the feature_engineer.py file. Different statistical model can be created and adjusted (in terms of parameters) in the models.py file. Different tasks (train, predict, CV, and etc.) can be added in the pipelines.py and called from the main.py file by creating a command function in the main.py file. settings.py file contains runtime global parameters such as data filename, features/models to use.