The Data:

The Customer data collected came in the form of the following features:

  1. id; int64
  2. Total Book Length (in minutes); float64
  3. Book Length Average (in minutes); int64
  4. Total Price of Books Purchased; float64
  5. Price Average of all Books Purchased; float64
  6. Left a Review (1 if yes, 0 if no); int64
  7. Review Score (0 to 10); float64
  8. Minutes listened; float64
  9. Minutes Completed; float64
  10. Support Request Made (1 if yes, 0 if no); int64
  11. Days Since Last Use Since First Purchase; int64
The Targets were collected 6-months after in the form of a boolean. 1 if the customer made another purchase within that time frame, or 0 if that customer didn't make any purchases.

Preprocessing/Cleaning

The main issue regarding cleaning the data was handling missing values. I found the best method was to impute all missing values within the Review Score feature with the average score. Doing so helped in two ways:

  • I didn't have to discard much of the dataset
  • It preserves the average of the app's Review Score feature

    • My Preprocessing mostly consisted of balancing the dataset. the majority (over 80%) of the targets were 0. I kept this same ratio for all test, validation, and test datasets.

    Modeling

    I used Tensorflow.keras to create my model. I also created a grid search to determine the optimal hidden layer size. the table on the right shows that a hidden layer size of 350 was the most optimal. Since all these models perform relatively well (over 90%), the main take away is that a hidden layer size of 350 not only gives us the best performing model but does so at the lowest computational cost.

    Conclusion

    Given our data, we can correctly predict whether 9 out of 10 individuals will continue to use this Audiobook platform.