EDA & Classification on Kaggle Hotel Booking dataset
In this notebook, I did an exploratory data analysis and a classification using Random Forest model on Kaggle's Hotel Booking dataset. Specifically, I
- analyzed the general shape and columns of the dataset
- examined the 4 columns that contain missing values and if to impute or drop them in the dataset
- made univariate data analysis with 9 customer-related columns to acquire a general idea about the hotel customers, and
- evaluated 6 classification algorithms and predicted users' reservation cancellation chance using Random Forest model.
After tuning the parameters and modifying the required_car_parking_spaces column, the model achieves a cross-validation accuracy of 87.12%.
Here are some of the properties of the data I looked at for my EDA.
To effectivelly predict if a customer will cancel the booking, I looked at the columns that have the highest correlations with the cancellation column as well as the relationships between categorical features with the cancellation column.
When selecting features, to prevent possible data leakage, I excluded booking_changes, which may include the final cancellation of hotel reservation. Furthermore, reservation_status also may include whether a customer cancels the booking or not. So I also omitted that one. Thus, from the analysis, I selected these variables as input.
Then I implmented six classification algorithms using these features as input and checked the models' cross-validation accuracy with the dataset. From the results I decided to use Random Forest classifier for this task as it has the highest baseline accuracy.
For each column with high correlation, I visualized the association between the column and the cancellation column. Specifically, I found that if the customer requires at least one parking space, the chance to cancel the reservation is minimal:
Therefore, I decide to transform this column into a two-classes column that only has 0(which means the customer doesn't require parking space) and 1(which means the customer requires parking space). The modification is as follows:
Before Modification:
After Modification:
With this modification, the max and min score of the model both increased slightly:
Finally, after tuning the parameters, the performance of the model is about 87.12%.
