Hotel Booking Cancellation Prediction

Data Science project for analyzing and visualizing hotel booking dataset and predicting booking status.

Link to Kaggle Project Link to Kaggle Notebook here

Installation

Make sure you have Python installed. Clone the repository and install the required dependencies.

git clone https://github.com/AslihanYoldas/ada-lovelace-final-project.git

cd ada-lovelace-final-project.git

pip install -r requirements.txt 

Usage

Run the run.py script to load dataset, data preprocessing, feauture selection, training models and visualize evulation of the models. In module folder with three module files the functions that are used in the project and the used libraries defined.

For seeing the detailed steps through projcet you can look at the notebooks. If you want to see visualization of the data yo can look at the data viz notebook. The steps are explained below with the notebook links.

python run.py

Data

Dataset link

• Dataset contains 36k rows and 27 columns.
• Original Columns:
  • Booking_ID string - ID of the booking (unique)
  • number_of_adults int - number of adults
  • number_of_children int - number of children
  • number_of_weekend_nights int - number of weekend nights
  • number_of_week_nights int - number of week nights
  • type_of_meal category - meal type (Meal type 1, Meal Type 2, Meal Type 3, Not Selected)
  • car_parking_space int- car parking space (0,1)
  • room_type category - room type (Room Type 1, Room Type 2, Room Type 3, Room Type 4, Room Type 5, Room Type 6, Room Type 7)
  • lead_time int - number of days between the booking date and the arrival date
  • market_segment_type category - market segment type (Offline, Online, Corporate, Aviation, Complementary)
  • repeated int - Indicates whether the booking is a repeat booking (0,1)
  • P_C int - number of previous bookings that were canceled by the customer prior to the current booking
  • P_not_C int - number of previous bookings not canceled by the customer prior to the current booking
  • average_price float - average price of the booking
  • special_requests int - number of the special request
  • date_of_reservation datetime - Date of the booking
  • booking_status str- booking status (Canceled, Not Canceled)
• New columns : This columns created by using existing columns
  • reservation_day int - day of the reservation (1-31)
  • reservation_month int - month of the reservation (1-12)
  • reservation_year int - year of the reservation (2015-2018)
  • date_of_arrival datetime - date of the arrival
  • arrival_day int - day of the arrival (1-31)
  • arrival_month int - month of the arrival (1-12)
  • arrival_year int - year of the arrival (2016-2020)
  • lead_month int - number of months between the booking date and the arrival date
  • number_of_total_nights int - sum of week and weekend nights
  • number_of_total_people int - sum of adult and children

Data Preprocessing

Preprocessing-notebook

• Dataset doesn't have null values.

Handling Outliers

How it detected ?

For each column thresholds determined. The low limit is the 10th percentile minus 1.5 times the range. The high limit is the 90th percentile sum 1.5 times the range. Also scatter plots created for each column. Sometimes looking to the scatter plots new thresholds determined and sometimes outliers that calculated doesn't determined as outliers. For example with the case of number of adults the algorithm detected 4 adults as outliers. Because 4 adults doesn't seem to be extreme situation detected outliers left in the dataset.

How it handled ?

For each outlier some other columns values compared to other data. The rows that contain some values as the outliers found. In that rows calculated the median of the that outliers' column and replace outlier with it.

Outliers detected and replaced in:

• Average Price
• Number of childeren
• Number of week nights
• Number of weekend nights
• P-C
• P-not-C
  
  

Also because there is only one row has the arrival date of 2016 and three rows has the arrival date of 2020 that rows dropped. Because of droopping these rows reservation year 2015 and 2016 also dropped.

Data Visualazation

Data-Viz-Notebook

• Classes are imbalanced. Most of the data is classified not cancelled.

• Average Price Distribution group by booking status

---

• Lead Time Distribution group by booking status
• Most of the time there are 0-2 months between reservation and arrival
• Bigger the time difference between reservation arrival date bigger rate of the cancelation

• Special Request Distribution group by booking status
• Most of the time there are 0-2 special request
• More special request made lesser cancelation

• Mostly arrival to the hotel is in 2018 but there are also 207 and 2019 data

---

• In 2017 and 2018 Not Cancelled ratio bigger than Cancelled ratio but 2019 Canceled is more dominant

• There is a big increase in lead time in the bookings that have arrival date of 2019
• Because of lead time increased cancelation in 2019 is bigger than other years

---

• Corelation Heat Map
• lead time , market segment type, average price, special request,reservation year, arrival year and lead month has a correlation with booking status
• There is not a strong corelation with other variables except the obvious ones.

Modelling

Modelling notebook

Hotel booking cancelation prediction is a classifying problem. For this task Logistic Regression model, K-Nearest Neighbour Classifier, Decision Tree Classifier, Support Vector Classifier trained and evaluated.

Feauture Selection

Logistic Regression

For logistic regression a model trained with all the data and coefficents for every variable compared. Both positive and negative values higher than 0.5 are selected which are :

• special_requests
• average_price
• lead_time
• market_segment_type

Decision Tree

Similar to logistic regression a model trained with all tthe data and model's importance score for each variable compared. Values higher than the 0.05 are selected which are :

• lead_time
• average_price
• special_requests
• market_segment_type
• reservation day
• arrival day

KNN and SVC

For KNN and SVC select k best with f_classif score function and Extra Tree Classifier models feaure importance score used. Both techiques returned the same features which are :

• lead_time
• average_price
• special_requests
• market_segment_type
• reservation day
• reservation month
• arrival day
• lead month
• arrival month

Model Training

Dataset normalized and for each model spillited 80/20 ratio using selected features.

Model Evaluation

Model Evaluation made with confusion matrix and classification reports.

• Logistic Regression

• KNN

• Decision Tree

• SVC

---

F1 Score Table

Model	Cancelled	Not Canceled
Logistic Regression	0.66	0.85
KNN	0.75	0.89
Decision Tree	0.77	0.88
SVC	0.68	0.87

Results

• Not canceled class prediction is better than cancelled. It is most likely because of imbalance of classes.
• Best canceled prediction made with decision tree at %77
• Best not canceled prediction made with knn at %89
• KNN and Decision Tree are the best resulted models.

Contributing

Feel free to contribute by opening issues or submitting pull requests.

Contact

For questions or feedback, contact aslihanyoldas24@gmail.com