Classification of hotel guests by predicted additional spending with ANN decision support system

Abstract

© 2017 IEEE. This paper presents a decision support system for classification of hotel guests in the terms of additional spending. The research is conducted on three stars medium-sized hotel. Guests are classified on arrival, during check-in, in one of the two groups: low spending group or high spending group. A low spending group consists of visitors that are anticipated to spend less than 25 Euros per day for additional hotel services. Contrary, a high spending group consists of visitors that are anticipated to spend more than 25 Euros per day on additional spending. The purpose of the research is to design a decision support system to predict an average daily spending of a guest based on available check-in information. The marketing department of a hotel can exploit this information (if available) and adapt promotions of specific goods and services, provided in the hotel, to meet specific customers' needs. This personalization of hotel promotions are expected to increase income, reduce costs and improve the overall image of a hotel in customer ratings. The input parameters of a classifier are derived from the following: how many days in advance a booking is made; how long a visitor plans to stay in the hotel; the price of a daily arrangement and the country of origin. The county of origin is numerically presented by three statistical parameters: GINI coefficient, HDI (Human Development Index) and GDP (Gross Domestic Product) per capita. Artificial neural network classifier is proposed since observed feature space is six dimensional and nonlinear. For classifier selection a new criteria is proposed as a minimum distance from an ideal classifier in receiver operating characteristic plot. Proposed measure is simpler for calculation than Matthew's correlation coefficient and gives information of the overall performance of the classifier. The proposed classifier proved a performance of 84% of correctly classified guests on test data set, which is quite satisfying result for this kind of application.