This notebook serves as an improvement of the training project of hotel booking cancellation prediction. This model use more data and features from the dataset, with refined background and more explanation at each notebook's section.
Let's say that we have a hotel business to run. Since it is a business that revolve around service, customer satisfaction is the most important factor that indicates our success and future prospect of the business. Good experience will lead to good reviews, it will help us in developing loyal customer and searching for new customers. Increase in customers drives most businesses, it will lead to increase in revenue and business growth, that we can increase the quality of service and also expand our business. It is basically a snowball effect.
First impressions are crucial in business because it's a human nature to make judgment about someone in their first meet. Making a strong first impression will help the business develop customer relationships and sales. In hotel business, preparing for the arrival of the customers is important to develop a strong first impression. However, how does the hotel management distinguish between customers that will come or not based on their characteristics in booking application?
In this project, I will aid the hotel management to predict which customers that will cancel their booking based on their booking application characteristics. With this predictive machine learning model, it will lead to cost-efficiency of the hotel management in preparing the customers arrival and increase in customers satisfaction based on the strong first impression provided.
This is a real-world data of customers, obtained from 2 hotels which are located in Portugal: Resort Hotel at the resort region of Algarve and City Hotel at the city of Lisbon. You can download it from Kaggle. The dataset initially consists of 119,390 rows with 32 columns.
Detailed data cleaning and EDA can be read at Jupyter. The process of data cleaning eliminates 0.4% of data into 118,902 rows and only use 19 columns (1 label+18 potential features).
From the graph above, we can see the imbalance of the dataset. Considering the amount of 0 class is higher than class 1, we can use undersampling method such as Random Undersampler or Near Miss to balance the dataset for more unbiased result.
Using hypothesis testing procedure (Mann-Whitney U), there was a signifant difference in median of days_in_waiting_list between 0 and 1 class, where class 0 (check-in) has longer duration of waiting. Thus, longer waiting duration didn't affect the booking cancellation. However, from this information, it is a precaution toward hotel owner to be able to give the best impression to those who waits longer. If failed, then it can easily turn into bad reviews for the hotel.
Which countries were planning to visit the hotel? Which countries has the highest cancellation rate?
We can observe top countries who visited the hotel from the graph above. From the heatmap, country with the highest cancellation rate is Portugal. Irony, isn't it? Well, logically, since Portugal is the nearest to the hotel, they can cancel and booking for another time. Maybe we can make a new feature based on Region of the country that will indirectly indicates the distance to the hotel.
Interestingly, most of the non-refund deposit type were resulting in booking cancellation. That's probably because it consists of deposit with a full cost of stay which really put the customers at huge disadvantage/risk of losing money. This feature will probably dominate the prediction model. This information can be a suggestion to the hotel management, because other than predicting the booking cancellation, one other thing that interesting to be answered is how to reduce the booking cancellation. To minimalize the booking cancellation rate, the hotel management needs to alter this deposit type into another kind that put both hotel owner and customer advantages in mind.
Room A is unquestionably the most popular room. Based on the heat map above, room H has the highest cancellation rate, followed by A and G. Room P cancellation didn't represent the group because it only consists of 2 observation.
To select the appropriate evaluation metric, we need to understand about the risk of false negative and false positive in our prediction case. In short term point-of-view (POV), false negative of booking cancellation prediction is more likely resulting in profit-loss because of unnecessary preparation of the hotel towards non-existent customer. But in the long term POV, failing to prepare for the customer's arrival also will reduce our customer's experience, resulting in unseen profit/opportunity-loss in the future (false positive). So, balancing false positive and false negative (F1 score) is necessity for this prediction model building.
Based on its unique value for categorical data and range for numerical data, we will do the following:
1. Robust scaling for numerical data with broad range (with outliers);
2. Binary encoding for categorical data with lots of unique values;
3. One hot encoding for categorical data with few of unique values.
Considering the mean score and its standard deviation, XGBClassifier was used.
A good F1 score also observed in the test set, indicating a good fit/generalization. Can we improve the model using hyperparameter tuning of the machine learning method?
Hyperparameter tuning didn't improved the model quality. To improve the model, we need to increase the quality of the data, by removing outliers, anomalies, etc.
This result confirms our argumentation in the EDA process of deposit type. This information can help us to understand the main factors that contribute to the booking cancellation, also in determining strategies on how to reduce the booking cancellation rate. One thing for sure, the hotel management needs to modify the non-refund system of deposit, maybe with only 50% of total cost of stay.
Compared to other analysis of the same dataset at Kaggle, you can see that our F1 score is better than most of the models. I believe that the process of feature elimination based on its definition, using logic really helps. The process of undersampling also balanced the data, thus improving the F1 score compared to other models in Kaggle. Hyperparameter tuning of the model didn't improve the model, thus to further improve the model, we need to improve the quality of the data. Overall, our model can predict the hotel booking cancellation quite well with F1 score of 75% in test set, with the most importance factors are deposit type (non-refund), car parking space, region that indicates distance, market segment, and also previous cancellation history.