Fraudulent mobile ad clicks have become an increasingly difficult problem for advertisers as the world experiences a proliferation of smart mobile devices. Advertisers pay affiliate websites when their users click on one of their mobile ads.
This compensation model has incentivized bad actors to generate fraudulent clicks on mobile ads in the hope of conning advertisers into paying them a commission. Advertisers must compensate affiliates even when these clicks do not result in a download of the advertised application.
In response to this problem, we've developed a machine learning model that can predict whether a user will download an app after clicking a mobile advertisement. Training our model on a dataset of around 100,000 mobile advertisement clicks spanning three days, we were able to achieve 83% accuracy using XGBoost.
This model can be used by advertisers to segregate fraudulent clicks from genuine ones. This distinction saves advertisers significant money and thwarts bad actors.
- main.ipynb: The Jupyter Notebook where the source code for our models is written and executed.
- eda.py: The Jupyter Notebook for exploratory data analysis on our data sets.
- equalizer.py: The script that parses the original 184 million training set and returns an equalized and downsized CSV of training data.
- Dockerfile: Configuration and setup of the container that the notebook will be running in.
- Pipfile: Configuration for the project Pipenv virtual environment. Used to install specific project dependencies and version of Python.
- Pipfile.lock: Contains meta-data for the Pipenv virtual environment.
- __init__.py: Make the project root a Python module.
- data/: Directory with all original and engineered training data
- data/train.csv: 184,903,890 rows of the original training data provided by Kaggle. (this is omitted from the submission because of the 8 GB file size, please download from google drive link)
- data/train_sample.csv: 100,000 randomly-selected rows of training data provided by Kaggle
- data/equalized_train.csv: 91,473 rows of equalized and randomly selected training data.
- data/test.csv: 18,790,469 rows of test data provided by Kaggle to evaluate our models with.
- output/: Directory with prediction files
- output/prediction_model_acc.csv: file pattern for the predictions and accuracy of the best models (dtc, knn, randomforest, and xgboost)
- .gitignore: Includes file patterns not to be included in Git.
- .dockerignore: Includes file patterns not to be added to the Docker image.
- README.md: Project description and direction on how to run the model.
- CMPS_144_FINAL_PAPER.pdf: Final Project Report
This project has multiple dependencies and only supports Python 3.6. Since each of us work on different operation systems, we decided to run the project inside a Docker container along with Pipenv. This allows our project to run the same and automate installation regardless of the host OS. An up to date installation of Docker is required to complete the following steps.
Step 1: Build the Docker Image:
docker build -t fraud_detection .
Step 2: Run the Docker Container:
docker run -p 8080:8888 -v $(pwd):/fraud_detection fraud_detection
Step 3: Open the Jupyter Notebook in Your Browser
Once the Docker container is running, you can go to http://localhost:8080 to view the notebook. When prompted, enter the password/token found in the output of your running Docker container.
You can run the project notebook directly by running jupyter notebook main.ipynb. This is not recommended since you might not have the correct version of the project's dependencies.
Google Drive (only viewable by team members, but project submisison contains all relevant files)
Kaggle Challenge