Deep neural network-based tools for hotel marketing activities of online travel agents

Abstract Vacationers always want to choose the hotel that best meets their needs and expectations. For this, they typically use online travel agents (OTAs). OTAs must analyze millions of images from thousands of hotels quickly and classify them accurately to best meet demands. This is a highly complex and challenging task. OTAs also have to choose a default image to represent best way each hotel in order to best market hotels to vacationers and encourage click behavior. This is another challenge to overcome. The biggest problem in classifying millions of images from different areas of hotels is that these areas often have similar objects and environments. For this reason, fewer categories have been used in deep learning (DL) studies conducted so far. In this study, two deep neural networks (DNN) have been developed to accomplish these challenging tasks. The first network is the "hotel image classification network" (HICNet), which classifies images of different areas of hotels into 11 categories. In a multi-class classification task that includes closely related categories such as bar, food, and lobby, the HICNet network achieved very good results, reaching an average of 79% precision, sensitivity, and F1-score. The second network is the "hotel default image network" (HDINet), which enables the determination of a default image for each hotel on the OTA page. HDINet can accurately identify the default image of hotels with an average of 92% precision, sensitivity, and F1-score. These networks have been integrated into the OTA system named enuygun.com and are currently in active use.