Adaptive tourism forecasting using hybrid artificial intelligence model: a case study of Xi'an international tourist arrivals

Shuxin Zheng, Zhongguo ZhangPeerJ Comput. Sci 2023 · DOI

TL;DR

This paper introduces a novel hybrid GM-LSTM model to improve tourism demand forecasting, particularly for datasets with limited historical data. By combining the trend-capturing ability of the Gray Model (GM) with the non-linear pattern recognition of Long Short-Term Memory (LSTM) networks, the model achieves higher accuracy and efficiency compared to traditional methods.

What Problem Does This Paper Address?

Accurate tourism demand forecasting is crucial for businesses and policymakers. However, many advanced deep learning models require extensive datasets, which are often unavailable for tourism forecasting due to insufficient historical data or the complexity of the factors influencing travel. Existing time series models struggle with non-linear behaviors and market volatility, while traditional AI models can be data-hungry. This paper aims to bridge this gap by developing a forecasting model that performs well even with small, non-linear datasets.

Methods

The core of this research is the GM-LSTM hybrid model. It works in two main stages:

  1. Gray Model (GM(1,1)) for Trend: The first-order Gray Model (GM(1,1)) is used to capture the overall trend of the tourism data. This model is known for its effectiveness with limited data.
  2. LSTM for Residual Fluctuations: After the GM model extracts the trend, the remaining residual fluctuations (errors) are then modeled and predicted using a Long Short-Term Memory (LSTM) neural network. LSTMs are well-suited for capturing complex, non-linear patterns in sequential data.

A rolling window scheme is employed for both models. This means that instead of using the entire historical dataset for training at once, the models are trained on a sliding window of recent data. This allows the model to adapt to changing trends and capture real-time system characteristics. The optimal window size was determined through a grid search based on MAPE.

For comparison, the study also implements and evaluates a standard ARIMA(2,2,4) model, a common time series forecasting technique.

Datasets

The study uses historical data on annual international tourist arrivals in Xi'an, China, from 1980 to 2018.

  • Training Set: 1980-2013
  • Test Set: 2014-2018

Key Insights and Ideas

  • Hybridization is Key for Small Data: The GM-LSTM model demonstrates that combining a model-based approach (GM) with a data-driven approach (LSTM) can overcome the limitations of each when dealing with sparse data. The GM handles the overall trend effectively, while LSTM refines predictions by capturing the intricate fluctuations.
  • Rolling Window for Adaptability: The use of a rolling window is crucial for making the GM model more adaptive and improving its forecasting accuracy, as it prevents the model from being overly reliant on potentially outdated historical data.
  • LSTM Alone Can Overfit: A standalone LSTM model, while performing well on the training data, showed a significant drop in performance on the test set, indicating overfitting issues when training data is limited. This highlights the benefit of the GM component in providing a stable trend baseline.

Key Results

The GM-LSTM model achieved a predicted MAPE of 11.88% on the test set, outperforming other models.

  • GM-LSTM: MAPE of 10.24% on the test set.
  • Rolling GM: MAPE of 11.88% on the test set.
  • ARIMA(2,2,4): MAPE of 19.31% on the test set.
  • Standalone LSTM: MAPE of 22.33% on the test set.

The GM-LSTM model not only showed superior accuracy but also captured an enhanced growth trend in the data.

Limitations and Future Work

The authors acknowledge several limitations and suggest avenues for future research:

  • Univariate Data: The current model only uses monthly tourist arrival data. Future work should incorporate multivariate inputs such as GDP, population, and seasonality to enhance accuracy and provide deeper insights.
  • Hyperparameter Configuration: Model hyperparameters and architecture were manually configured. Exploring automated methods like neural architecture search or Bayesian optimization could lead to better configurations.
  • Lack of Prediction Intervals: The model currently provides point estimates. Developing a probabilistic version of GM-LSTM could generate predictive distributions, which would be more useful for quantifying uncertainty.
  • Interpretability: As a "black-box" AI model, the GM-LSTM's decision-making process is not transparent. Future research should focus on techniques for improving model interpretability and transparency.

Summary updated

BibTeX
@article{zheng2023adaptive,
  author = {Shuxin Zheng and Zhongguo Zhang},
  title = {{Adaptive tourism forecasting using hybrid artificial intelligence model: a case study of Xi'an international tourist arrivals}},
  journal = {PeerJ Comput. Sci.},
  year = {2023},
  doi = {10.7717/peerj-cs.1573},
}
Figure