Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A novel forecasting framework combining virtual samples and enhanced Transformer models for tourism demand forecasting

Tingting Diao, Xinzhang Wu, Lina Yang, Ling Xiao, Yunxuan Dong

TL;DR

This paper tackles forecasting tourism demand under limited historical data and complex spatiotemporal dependencies. It introduces a framework that fuses a spatiotemporal GAN for adaptive virtual sample generation with an enhanced Transformer predictor featuring causal convolution and non-autoregressive global pooling. The approach includes a rolling dynamic spatial weight matrix learned by the GAN and a joint training scheme where predictor feedback guides virtual sample generation. Empirical results on Macau and Turkey datasets show substantial improvements in forecast accuracy, with an 18.37% reduction in average MASE compared to conventional Transformer models, demonstrating robustness to data scarcity and improved capture of spatiotemporal patterns. The work offers a practical, scalable tool for tourism management and provides a blueprint for applying adaptive spatiotemporal augmentation to other domains.

Abstract

Accurate tourism demand forecasting is hindered by limited historical data and complex spatiotemporal dependencies among tourist origins. A novel forecasting framework integrating virtual sample generation and a novel Transformer predictor addresses constraints arising from restricted data availability. A spatiotemporal GAN produces realistic virtual samples by dynamically modeling spatial correlations through a graph convolutional network, and an enhanced Transformer captures local patterns with causal convolutions and long-term dependencies with self-attention,eliminating autoregressive decoding. A joint training strategy refines virtual sample generation based on predictor feedback to maintain robust performance under data-scarce conditions. Experimental evaluations on real-world daily and monthly tourism demand datasets indicate a reduction in average MASE by 18.37% compared to conventional Transformer-based models, demonstrating improved forecasting accuracy. The integration of adaptive spatiotemporal sample augmentation with a specialized Transformer can effectively address limited-data forecasting scenarios in tourism management.

A novel forecasting framework combining virtual samples and enhanced Transformer models for tourism demand forecasting

TL;DR

This paper tackles forecasting tourism demand under limited historical data and complex spatiotemporal dependencies. It introduces a framework that fuses a spatiotemporal GAN for adaptive virtual sample generation with an enhanced Transformer predictor featuring causal convolution and non-autoregressive global pooling. The approach includes a rolling dynamic spatial weight matrix learned by the GAN and a joint training scheme where predictor feedback guides virtual sample generation. Empirical results on Macau and Turkey datasets show substantial improvements in forecast accuracy, with an 18.37% reduction in average MASE compared to conventional Transformer models, demonstrating robustness to data scarcity and improved capture of spatiotemporal patterns. The work offers a practical, scalable tool for tourism management and provides a blueprint for applying adaptive spatiotemporal augmentation to other domains.

Abstract

Accurate tourism demand forecasting is hindered by limited historical data and complex spatiotemporal dependencies among tourist origins. A novel forecasting framework integrating virtual sample generation and a novel Transformer predictor addresses constraints arising from restricted data availability. A spatiotemporal GAN produces realistic virtual samples by dynamically modeling spatial correlations through a graph convolutional network, and an enhanced Transformer captures local patterns with causal convolutions and long-term dependencies with self-attention,eliminating autoregressive decoding. A joint training strategy refines virtual sample generation based on predictor feedback to maintain robust performance under data-scarce conditions. Experimental evaluations on real-world daily and monthly tourism demand datasets indicate a reduction in average MASE by 18.37% compared to conventional Transformer-based models, demonstrating improved forecasting accuracy. The integration of adaptive spatiotemporal sample augmentation with a specialized Transformer can effectively address limited-data forecasting scenarios in tourism management.

Paper Structure

This paper contains 26 sections, 24 equations, 5 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Literature review
  3. Spatiotemporal tourism demand forecasting
  4. The GANs for virtual sample generation
  5. Transformer-based models as predictor
  6. Proposed method
  7. Virtual sample generation module for forecasting framework
  8. Predictor module for forecasting framework
  9. Model overview
  10. Causal convolution for local feature extraction
  11. Positional encoding and transformer encoder
  12. Global pooling and prediction head
  13. Joint training predictor and virtual sample generation module
  14. Empirical study
  15. Sample analysis
  16. ...and 11 more sections

Figures (5)

  • Figure 1: The proposed predictor module for forecasting framework.
  • Figure 2: The joint training of the proposed predictor. The strategy is divided into two parts: (1) Pre-training the predictor model. (2) Training the spatiotemporal GAN and predictor model.
  • Figure 3: Heat map of correlation among countries or regions (The purple border highlights strong correlations ($|r| \geq 0.5$), while the blue border indicates moderate correlations ($0.3 \leq |r| < 0.5$).).
  • Figure 4: Radar chart of MAPE for tourism demand using various models (Macau dataset: purple, blue, green, and pink were used to fill the MAPE forecasting horizons of 1, 3, 5, and 14 steps).
  • Figure 5: Radar chart of MAPE for tourism demand using various models (Turkey dataset: purple, blue, green, and pink were used to fill the MAPE forecasting horizons of 1,3, 6, and 12 steps).