The use of artificial neural network algorithms to enhance tourism economic efficiency under information and communication technology

Datasets collection

To validate the effectiveness of the proposed ANN model, this work collects various types of data from a smart tourism site as research samples. These data encompass tourists’ basic information, consumption records, satisfaction ratings, and more, providing a rich information source for model training and evaluation.

(1) Data Source: The data for this work are sourced from a well-known smart tourism site in China. This site has implemented a range of ICT technologies, including but not limited to mobile payment systems, visitor behavior tracking systems, and social media platforms, which provide a solid foundation for data collection.

(2) Data Types: The main data types involved here are as follows. Tourist Basic Information: Demographic characteristics such as age, gender, and occupation. Consumption Records: Information on tourists’ expenditure within the site, including the amount spent, time of expenditure, and location.

Satisfaction Ratings: Collected via online surveys, including ratings for services such as dining, accommodation, and attraction quality.

(3) Data Collection Methods: Online Surveys: Questionnaires are distributed via the site’s official website and social media platforms to gather tourists’ basic information and satisfaction ratings. Mobile Payment Systems: Data on tourists’ expenditures are obtained through third-party payment platforms partnered with the site. Visitor Behavior Tracking Systems: Location sensors and Wi-Fi hotspots installed at the site are used to collect information on visitors’ movement trajectories and dwell times.

(4) Data Preprocessing: Before using the data, preprocessing is necessary to ensure its quality and consistency. The preprocessing steps are as follows. Data Cleaning: Removing duplicate data and correcting format errors or inconsistencies. Missing Value Handling: Filling missing data with mean, median, or mode values, or predicting missing values using other variables. Outlier Detection: Identifying and addressing outliers through statistical analysis.

(5) Data Privacy Protection: During data collection, strict adherence to relevant laws and regulations is maintained to protect tourists’ privacy. Measures such as anonymizing sensitive information are implemented to ensure personal identity information remains confidential.

(6) Data Format: All collected data are converted to a standardized format to facilitate subsequent data processing and analysis. The main data formats are as follows. Structured Data: Stored in tabular form, such as Excel files. Unstructured Data: Includes text comments and images, which are converted through specialized processing workflows.

Experimental environment

To ensure the proposed ANN-based smart tourism model can be effectively trained and evaluated, experiments are conducted on a high-performance computing platform. Table 1 displays the hardware configuration.

This hardware configuration meets the requirements for large-scale data processing and complex model training, with GPU significantly accelerating the computation process, especially for deep learning tasks.

Table 2 presents the software environment.

Parameters setting

Hyperparameters are parameters that need to be set manually before training a model, and they control the behavior of the learning process^37,38,39. Proper hyperparameter tuning is crucial for the performance of the model. For example, in multiple experiments, 256 hidden units are found to provide sufficient model capacity to capture complex non-linear features while avoiding overfitting caused by an excessive number of hidden units. The experimental results show that using fewer hidden units (such as 128) leads to insufficient learning capability, while more units (such as 512) increase model complexity but do not significantly improve performance and add computational overhead. Regarding the choice of activation function, this work selects the ReLU function instead of Leaky ReLU or Swish. ReLU has shown good convergence when training deep neural networks and effectively avoids the vanishing gradient problem. Although Leaky ReLU can mitigate the “dead neuron” issue of ReLU and Swish performs excellently in certain tasks, ReLU provides the most stable experimental results in this task, and it is computationally efficient. Therefore, the selection of ReLU as the activation function was based on its best performance here. Table 3 presents the key hyperparameters considered in the model training process.

After determining the hyperparameters, this work uses cross-validation to obtain performance metrics for the model under different parameter combinations, including accuracy, precision, recall, and F1 score. By comparing these metrics, it ultimately selects the hyperparameter combination that achieves the highest average performance.

Performance evaluation

(1) Model Performance Evaluation Results. This work evaluates the performance of the proposed model, and Fig. 3 presents the results.

Model performance evaluation results.

Figure 3 suggests the performance variations of the model under different settings for learning rate, batch size, and optimizer. The model with a learning rate of 0.001, a batch size of 128, and the Adam optimizer performs the best across all metrics.

(2) Comparison with Traditional Statistical Methods. This work compares the proposed model with traditional models. Figure 4 displays the results.

Comparison with traditional statistical methods.

Figure 4 reveals that the ANN outperforms traditional statistical methods across all performance metrics, particularly in the F1 score, indicating that the model performs better in balancing precision and recall. Decision trees (DT) exhibit the weakest performance across all metrics, likely due to their relative simplicity and inability to capture complex relationships in the data. Random forest (RF) and support vector machine (SVM) have comparable performance but are slightly inferior to the ANN model in terms of accuracy and F1 score. When compared with XGBoost and Transformer, the ANN model still performs best across four metrics: accuracy, precision, recall, and F1 score, with an F1 score of 0.83, slightly higher than Transformer (0.82) and XGBoost (0.80). Although Transformer and XGBoost are close in some metrics, the overall performance of the ANN is more balanced, maintaining its lead across all indicators.

Additionally, a significance comparison between the ANN model and other models is conducted, and Table 4 shows the results.

According to the significance comparison results in Table 4, the ANN model significantly outperforms DT, RF, and SVM across the four metrics—accuracy, precision, recall, and F1 score—with p-values less than 0.05. This indicates that the ANN performs notably better than these traditional models. However, the differences between ANN and XGBoost or Transformer models in all metrics are not significant, as the p-values are greater than 0.05. This suggests that there is no substantial performance difference between these two models and ANN. Overall, while ANN shows a significant advantage when compared to traditional models, the difference is small when compared to advanced AI models.

To further understand the limitations of the model in practical applications, a detailed analysis of misclassification cases in the ANN model is provided below:

Case 1: Misprediction of Tourist Preference Changes. In some cases, the ANN model fails to accurately capture the rapid changes in tourist preferences. For instance, certain tourists behave differently during peak travel seasons than they do during normal time, and the model fails to effectively identify these abrupt changes, leading to incorrect predictions. Specifically, some tourists alter their attraction choices during peak periods due to factors such as weather or holidays, but the model does not adapt to this change in time, resulting in the misclassification of their behavior. This type of misprediction usually occurs in edge cases where tourist behavior has significant fluctuations, and the model’s generalization ability is somewhat limited in such scenarios.

Case 2: Misclassification of Low-Frequency Tourists. Another common type of misclassification occurs within the low-frequency tourist group, particularly for first-time visitors to a specific attraction. Due to the limited historical data for these tourists, the model struggles to gather enough information during training to accurately predict their behavior. For example, when predicting whether a tourist will choose a particular attraction, the model might incorrectly classify them as “not choosing,” whereas, in reality, the tourist may have chosen the attraction for a special reason (such as recommendations and social media influence). This behavior pattern of low-frequency tourists is often a weak point for the model, especially when sufficient sample data is lacking, which leads to a decrease in prediction accuracy.

Case 3: Noise Data Impact. In practical applications, noise data also affects the model’s prediction results. For instance, some tourists may input incorrect basic information or consumption records due to operational errors or misunderstandings when filling out surveys. These noise data points are not effectively identified or filtered, which leads to incorrect predictions by the model. During the data preprocessing stage, efforts are made to remove some noise data. However, due to the inherent imperfection of real-world data, some misleading or inconsistent samples still remain, and these samples can affect the prediction performance of the ANN model.

Through the analysis of the above misclassification cases, it can be observed that the ANN model tends to exhibit prediction biases when faced with specific scenarios, such as rapid changes in tourist preferences, behaviors of low-frequency visitors, and interference from noise data. The root cause of these misclassifications lies mainly in the marginal characteristics of the data or the presence of noise, which prevents the model from accurately learning all possible behavior patterns. Therefore, to further improve the accuracy of the model, future research can delve into areas such as data augmentation, noise data handling, and optimization for low-frequency groups.

To ensure the practical feasibility of the ANN model, especially for deployment in the smart tourism field, its computational costs have been carefully evaluated. In practical applications, the computational overhead of the ANN model primarily stems from the time consumed during the training and inference phases, and the required hardware resources. During the model training phase, the computation time of the ANN is closely related to the size of the dataset, the complexity of the model, and the hardware performance. Taking the dataset used as an example, the training process takes about 8 h. This process includes steps such as data preprocessing, feature engineering, model optimization, and hyperparameter tuning. With a sample size of 100,000 tourist behavior records, the ANN model completes the training within a relatively reasonable time, demonstrating good scalability. Furthermore, the computational cost during the inference phase is relatively low. For behavior prediction of each tourist, the inference time for the ANN model is approximately 10 ms, meaning it can process data from about 100 tourists per second. This is completely feasible for real-time recommendation and resource scheduling applications. Although the ANN model incurs certain computational overhead, with reasonable hardware configuration and optimization, the model can meet the demands of large-scale real-world applications. Moreover, with improvements in hardware performance and the use of distributed computing frameworks, the computational costs of the ANN model will continue to decrease. Therefore, based on the current technology, the deployment of the ANN model in the smart tourism field is both feasible and scalable.

(3) User satisfaction survey results. This work analyzes the results from 300 survey responses. Figure 5 displays the user satisfaction results.

User satisfaction survey results.

Figure 5 shows that most users express satisfaction or high satisfaction with the personalized service recommendations and resource optimization configurations, indicating that the model has been well received in practical applications. 34% of users are very satisfied with the personalized service recommendations, indicating that the model effectively meets individual user preferences. Regarding resource optimization and allocation, 76% of users express satisfaction or high satisfaction. This suggests that the optimization of resource distribution has received positive feedback from most users. For overall user experience satisfaction, 35% of users are very satisfied, which is slightly higher than the satisfaction with personalized service recommendations. This indicates that, in addition to the service recommendations, the model has also improved the overall user experience.

(4) User Feedback Results. This work analyzes the results from the survey responses. Figure 6 presents the user feedback results.

Figure 6 shows that most users believe the recommended services align well with their personal preferences, with high proportions agreeing or strongly agreeing. This indicates that the model effectively captures user preferences. Users generally feel that service quality has improved, suggesting a positive impact from the model’s application. Regarding the reasonableness of resource allocation, while there is a high agreement, the proportion of those who strongly agree is relatively lower, indicating room for improvement despite overall approval of the resource allocation. Most users agree or strongly agree that their overall experience has improved, highlighting the model’s effectiveness in enhancing user experience.

(5) User Suggestions. This work analyzes the survey responses and presents the user suggestions in Fig. 7.

User suggestions results.

Figure 7 reveals that most users consider it very important or important to add more services, indicating that users expect the model to offer a wider range of services. Regarding the need to improve recommendation accuracy, the majority of users believe this is important, suggesting they want the model to recommend services more accurately aligned with their interests. The importance of optimizing resource allocation is recognized by most users, indicating they wish for more effective use of resources in the tourism area. Users generally believe that enhancing the overall user experience is very important, suggesting they expect a smoother and more enjoyable travel experience. For strengthening privacy protection, although a higher proportion of users consider it important, the proportion who regard it as very important is relatively lower. This suggests that while users are concerned about privacy protection, they may be willing to compromise to some extent for a better service experience.