Fraudulent-Click-Prediction-in-Online-Advertising

For companies that advertise online, click fraud can happen at an overwhelming volume, resulting in misleading click data and wasted money. We build an algorithm that predicts whether a user will download an app after clicking a mobile app ad.

Dataset

• Size: 1441685 rows x 7 columns
• Attributes:
  • ‘ip’: ip address of click
  • ‘app’: app id for marketing
  • ‘device’: device type of user mobile phone
  • ‘os’: os version of user mobile phone
  • ‘channel’: channel id of ad publisher
  • ‘is_attributed’: indicates if the app was downloaded. (Target)
• Source: Kaggle competition (TalkingData Ad Tracking Fraud Detection Challenge)
• Class imbalance: 99.65% : 0.35% (0 : 1)

Imbalance Learning Strategy Employed

• Under Sampling: Only a very small part of the is_attributed data have 1 value (0.35%). This means that the training dataset is highly imbalanced . Usually either we undersample the records with is_attributed = 0 or oversample records with is_attributed = 1. Because is a large dataset, it is a good option to do undersampling of records with is_attributed = 0.
• LightGBM: Classifier used to handle class imbalances.

Metrics

• F1-score
• Area under ROC curve