Click Through Prediction

Will they click our ad?

Executive Summary

Through exploration of the data, it was found that certain hours of the day and certain days of the week have a relationship with click-throughs. Also, the anonymized continuous features were seen to have different means when someone clicked and when someone didn't click so those were put into my model. I ran several different models and found that Random Forest could beat baseline. Until I can look more into other features, we can use this model to predict the click-through rate of future advertisements.

Project Description

In online advertising, click-through rate (CTR) is a very important metric for evaluating ad performance. In this project I will try to predict if someone will click-through based on information provided from the website.

Project Goals

• Identify key features that can be used to create an effective predictive model.
• Use classification models to make click-through predictions.
• Use findings to make recommendations and establish a foundation for future work to improve model's performance.

Initial Thoughts

My initial hypothesis is that those who are browsing the website in the middle of the night are more likely to click the advertisement.

The Plan

• Aqcuire the data from here

• Prepare data

  • Checked for nulls, there were none
  • Changed the hour column to datetime
  • Created day_of_week and hour_of_day based off of hour

• Explore data in search of drivers of churn

  • Answer the following initial questions
    • What percentage of instances result in a click-through?
    • Does hour of the day have a relationship with clicks?
    • Does day of the week have a relationship with clicks?
    • Does banner position have a relationship with clicks?

• Develop a model to predict the value of a house

  • Use drivers identified through exploration to build different predictive models
  • Evaluate models on train and validate data
  • Select best model based on highest accuracy
  • Evaluate the best model on the test data

• Draw conclusions

Data dictionary

Feature	Definition	Type
id	Advertisement instance unique id	int
hour	The date and hour of day	datetime
C1	Anonymized categorical variable	int
banner_pos	Location of ad on the page	int
hour_of_day	Hour of the day when the ad was displayed	str
day_of_week	Day of the week when the ad was displayed	str
C14-C21	Anonymized continuous variables	int
Target variable
click	Did they click? (1-Yes, 0-No)	int

Steps to Reproduce

1. Clone this repo
2. Run final_project.ipynb notebook.

Conclusion

Summary

• hour_of_day (0, 1, 8, 9, 11, 14, 15, 16, 18, 19, 20, 21) probably have a relationship withclicks.
• Each day_of_week probably has a relationship with click.
• banner_pos (0, 1, 2, 7) probably have a relationship with click.
• Proabably a difference in means between each anonymized continuous feature (C14-C21) for those who click and don't click.

Recommendations

• We need to invest in some cloud computing in order to run statisical tests and models on such large data sets.
• Even though my best model only performs 0.2% better than baseline, we can use it for now as we continue to improve our model.

Next Steps

• In the next iteration:
  • Use cloud computing to look more into all the different features.
  • Try hashing some of the features to see if that will improve my models.