Article Preview
Top1. Introduction
Under the current Internet + model, the tourism field has also been developed rapidly. More and more users are pursuing a more efficient and fast style of work, so they are more willing to make travel plans and find relevant travel destinations through the online travel system, and they shared their photos and published some opinions and other information in the process of traveling on the Internet (Lyu et al., 2019). These will undoubtedly promote people's understanding of various tourist destinations, but it causes the system information overload due to the explosive growth of this type of data (Liao and Nong, 2021; Du, 2021). Faced with complex and huge travel information, it is difficult for users to quickly extract their favorite travel information from the system. At present, most of the functions of online travel systems only provide basic information retrieval, and the functions are too single to satisfy the user's data analysis and extraction functions before traveling.
Chen et al. (2020) simulated the subway station building evacuation design based on a deep neural network (DNN) model, which is compared with the convolutional neural network (CNN) model, the classification data set pre-training model, and the You Only Look Once (YOLO) algorithm to verify the accuracy and training speed of the model algorithm. Wang (2020) proposed a classification and processing method of tourism product information based on deep learning using the word embedding in the data preprocessing stage. The CNN is adopted to process user and travel service item review information, and the DNN is selected to process the necessary information of users and travel service items. The results show that the model can maintain an excellent accuracy of 64.2% when a personalized recommendation list for users is generated. Law et al. (2019) used deep learning methods to study the prediction framework of Macau's monthly visitor arrivals. The empirical results show that deep learning methods are significantly better than support vector regression (SVR) and artificial neural network (ANN) models. Feizollah et al. (2019) used deep learning algorithms to calculate and analyze Twitter sentiment in their research. The CNN, long and short-term memory neural networks (LSTM), and recurrent neural networks (RNN) are used to improve the prediction accuracy and build prediction models, so as to realize the sentiment calculation of Twitter in a specific topic. Shi et al. (2019) established the sentiment analysis experiments by analyzing the latest articles and techniques based on dictionaries, traditional machine learning, deep learning, and mixed sentiment analysis methods; and the most advanced results in different sentiment analysis experiments are obtained. Paolanti et al. (2021) established a social geographic data framework for deep learning to describe the spatial, temporal, and population tourist flow of this rural tourist area and its vast coastal areas. Four specially trained DNNs are used to recognize and evaluate emotions based on two words and two characters respectively. The rough data set is selected to reduce the dimensionality of the index, the number of neurons in the multi-layer structure of BPNN is optimized by QSIA, QSFOA, QPSO, and QGA, respectively, and the deep learning model is applied to establish the optimal neuron number prediction model under the three algorithms to predict the non-linear return rate of actual stocks. The results reveal that the QSFOA-BPNN model shows the highest prediction accuracy among all models. Fudholi et al. (2021) used cosine similarity to measure the similarity between a person’s picture and the gallery of a tourist destination through their label vector. An image classifier model run from a mobile user device through Tensorflow Lite is applied to infer tags. There are a total of 40 tags, covering local tourist destination categories, activities, and objects. The model uses the most advanced mobile deep learning architecture EfficientNet Lite for training. Th EfficientNet Lite is undertaken as the basic architecture for several experiments, and it is obtained that the accuracy rate is more than 85% on average.