Abstract: With the upgrade of tourism informationization and the rise of intelligent city construction, the Internet of Things, Cloud Computing and other information technologies have been widely used in the tourism industry, making the development model and technical framework of intelligent tourism become a hot issue. Aiming at the problems of insufficient preparation and inadequate reception capacity of scenic spots in recent years, this paper proposes to apply machine learning algorithm to predict the number of tourists, so as to make an early response. In this paper, the characteristics of the application of the number of tourists are analyzed. The fixed buffer kernel online gradient descent algorithm is used to predict the number of tourists, and the actual number of tourists is brought into the algorithm for experiments. Finally, the rationality of the experimental process and results is analyzed.
Introduction
Predicting tourist numbers is a critical challenge for modern scenic spot management. Inaccurate forecasts lead to significant operational issues, ranging from inadequate reception capacity during surges to wasted resources during lulls. This research addresses two primary hurdles in forecasting: the inherent non-linearity of tourist flow—a dynamic system influenced by weather, holidays, and marketing—and the "Curse of Kernelization."
While kernel-based machine learning models like Support Vector Machines (SVMs) are adept at handling non-linear relationships, standard online learning versions suffer from a linear growth in support vectors over time. In a continuous data stream, this leads to a memory explosion that eventually renders the system computationally unfeasible. This paper explores the application of Bounded Online Gradient Descent (BOGD) to mitigate these issues.
Methodology
The core of the proposed approach is the Bounded Online Gradient Descent (BOGD) algorithm. This method combines the power of kernel-based learning with a mechanism to maintain constant computational complexity.
Key technical features include:
- Kernel-Based Learning: Data is projected into a higher-dimensional space to capture complex, non-linear seasonal trends.
- Online Updates: The model updates its parameters incrementally as each new data point arrives, allowing for real-time adaptation to changing conditions.
- Fixed Buffer Mechanism: To solve the memory issue, the algorithm utilizes an intercept parameter ($B$) that defines a maximum capacity for support vectors. When the buffer is full, the algorithm prunes the support vector with the smallest coefficient, ensuring that memory usage remains stable regardless of the duration of the data stream.
Experimental Setup
The study utilized a real-world dataset from the Jiuzhaigou Scenic Area in China to evaluate the algorithm's performance. The dataset spanned from June 1, 2012, to June 1, 2017, comprising 1,827 daily data points. To improve training efficiency, the data was normalized to a $[0, 1]$ range using Min-Max scaling, though the authors note that this approach presents a challenge for true online learning where future maximum values are unknown.
Results and Discussion
The experimental results demonstrate that the BOGD framework is highly effective at identifying macro-level trends in tourist demand. The model successfully tracks massive surges during peak holiday seasons and the corresponding lulls during off-peak periods.
As shown in Figure 2, the predicted values (red line) maintain strong agreement with the actual values (blue line) regarding overall seasonal fluctuations. While the model struggles with high-frequency "single-day" precision due to local noise, its ability to capture the stochastic and time-varying nature of the data is evident.
Furthermore, the use of the fixed buffer ensured computational stability. The calculation process did not exhaust system resources over the 1,827-day period, proving the algorithm's viability for long-term deployment in "Smart Tourism" systems. This efficiency trade-off—discarding low-weight support vectors for constant computational cost—appears to be a pragmatic compromise for real-world applications.
Conclusion and Future Work
This preliminary study confirms that Bounded Online Gradient Descent is a scalable and effective tool for tourist number forecasting. However, several areas for improvement remain. Future work will focus on developing "on-the-fly" normalization techniques that do not rely on global dataset knowledge, which is inconsistent with the philosophy of online learning. Additionally, incorporating external features such as weather data and social media trends may improve the model's granularity and daily precision.
Summary updated