1.5: Datasets Used

We will use 4 separate datasets throughout this book for regression and classification problems. We will benchmark model performance and compare the performance of different techniques. These datasets are explained in the first four datasets from sections 1.5.1 to 1.5.3. We will also use 2 datasets for signal processing in chapter 11. These datasets are explained in sections 1.5.4 and 1.5.5.

1.5.1    Hotel Booking Demand Datasets

This dataset ^[2] has demand data for 2 hotels. Hotel H1 is a resort hotel that attracts customers who will like to stay at the hotel for recreation purposes. Hotel H2 is a city hotel, that people visit for business purposes.

We will use data from hotel H1 for developing classification models to predict the likelihood of cancellation for reservations. The objective of the classification model is to minimize losses from cancellations and increase profitability. Many customers cancel their reservations. This could mean a loss of revenue. If we know in advance which reservation is going to be canceled, we can try to resell the room, even though it is not canceled at the moment. However, it can lead to a situation when both guests arrive at the hotel and demand the room they have reserved. To avoid this situation, our modeling objective is to build a model which can predict cancellation with a high degree of accuracy. In other words, 'precision' will be our cost function. As a result, we might have a model which has a low recall, but high precision. We are fine with such a model, as long as it can predict cancellations with a low margin of error. Imagine a hypothetical model that has 95 precision and 30 recall. It simply means, for all predictions it makes for cancellations, 95 times out of 100 could actually be a cancellation. Although the precision is not the ideal 100, 95 precision is a good score for a statistician. A low recall of 30 means, out of all cancellations that happened, the model could only identify 30 percent of such cancellations. For a statistician or machine learning engineer, recall value of 30 is a less than ideal situation. However, for business owners, this is an acceptable situation. For a hotel revenue manager, recall value of 30 means that the hotel will be able to reduce their cancellation losses by 30 percent. This is still better than a situation where the hotel cannot save anything without the model. If not all, in many cases businesses desire a workable model. A perfect model which meets all criteria of theoretical statistics and theoretical machine learning is not always desired by real world businesses.

Data for hotel H2 will be used for developing a regression model to predict the total occupancy for rooms for future check-in dates. This dataset was not in the format where it could have been used for the regression model. The author has used his domain knowledge in the hotel industry to perform preprocessing and data cleaning to bring it into a usable format. Hotels have different types of customers. Two major groups are 'transient' and 'group'. Transient customers often seek short hotel stays, such as people who travel to different cities for business purposes, or other customers who want short-term stays. We will like to develop a model to predict total transient bookings at the check-in date level, at each booking window between 0 to 100 days before the date of check-in for the city hotel H2.

Booking day is the day when customers book the room for the check-in date in advance. For example, if you want to visit Hawaii on Christmas of 2022 and you are reserving the room on the 1st of October, 2022, then October 1st, 2022 will be the booking date and December 25th, 2022, the day of Christmas will be the check-in date. The number of days between these two dates is called lead time.

Sometimes customers book rooms, but they do not show up. We will remove these customers from our modeling problem, as these are very negligible. ADR is the average daily rate, for any check-in date it represents total revenue from bookings divided by total rooms sold. We will remove records for ADR which seem outliers, as we cannot get any explanation from the original authors about the nature of outliers.

We will use ADR between the bottom 5th percentile and the top 99.99 percentile. There were transactions for which no value was specified for the number of adults and children as guests. These will also be removed.

Some customers buy rooms for continuous stay. If a customer is reserving on the 1st of January for staying at the H2 hotel on the 7th and 8th of January, the lead time for the 7th of January will be 6 days and for the 8th of January will be 7 days.

We assume that one room can accommodate 2 adults and 2 kids in a hotel. If a transaction has 2 adults and 2 kids, it will be considered as 1 room. If a transaction has 1 adult and 2 kids, 2 adults, and 1 kid, or 1 adult and 1 kid, it will be considered as 1 room.

If there are more than 2 adults, the number of rooms will be calculated as the number of adults divided by 2. If the result is a decimal, it will be rounded to a higher number. For example, if there are 3 adults, 3 divided by 2 is 1.5. We will round 1.5 to a higher ceiling of 2. Hence concluding that 3 adults can stay in 2 rooms. Similarly, children are also considered for calculating the number of rooms.

1.5.2    Car Sales

This is sourced from Kaggle ^[3] and has information about the attributes of used cars and its price in India. It will be used for developing regression models. We will use the 'Car details v3.csv' file from version 3 of the dataset. The original file has details of 8128 cars. We have removed certain records which had inadequate data or missing data. These are mileage features for values of 0.0kmpl or left blank. Engine, torque, or seats are left blank. max_power as 0, 'bhp' or blank. km_driven as 1. After removing these observations, 7905 observations remain. Mileage, engine, and max_power were converted to numeric features after removing string suffixes.

1.5.3    Coupon Recommendation

In-vehicle coupon recommendation Data Set ^[4] studies the behavior of individuals and whether they will accept or not accept coupons. This dataset was generated through a survey. This survey describes different driving scenarios and asks the person if they will accept the coupon.

1.5.4    Raman Spectroscopy of Skimmed Milk Samples

This dataset ^[5] has a Matrix of quantitative whole spectrum analysis of 45 spectra on 21451 on skimmed milk samples. We will use this for discussing Raman spectra in chapter 11.

1.5.5    Beaver Body Temperatures

This dataset has body temperature of 2 beavers ^[6], measured every 10 minutes by telemetry. This has data for less than a day for each of the two beavers. We will discuss this in chapter 11 for filtering method.