Description
Context
This is real real-time bidding data that is used to predict if an advertiser should bid for a marketing slot e.g. a banner on a webpage. Explanatory variables are things like browser, operation system or time of the day the user is online, marketplace his identifiers were traded on earlier, etc. The column 'convert' is 1, when the person clicked on the ad, and 0 if this is not the case.
Content
Unfortunately, the data had to be anonymized, so you basically can't do a lot of feature engineering. I just applied PCA and kept 0.99 of the linear explanatory power. However, I think it's still really interesting data to just test your general algorithms on imbalanced data. ;)
Inspiration
Since it's heavily imbalanced data, it doesn't make sense to train for accuracy, but rather try to get obtain a good AUC, F1Score, MCC or recall rate, by cross-validating your data. It's interesting to compare different models (logistic regression, decision trees, svms, ...) over these metrics and see the impact that your split in train:test data has on the data.
It might be good strategy to follow these Tactics to combat imbalanced classes.