Tourist arrival forecasting with game-theoretic social influence weighting and deep learning

arXiv:2507.03411 · cs.LG, cs.GT, eess.SP · Submitted 2025-07-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: Today's paper: "Tourist arrival forecasting with game-theoretic social influence weighting and deep learning".

Jane: The gist The proposed framework integrates game theory-based opinion leader detection, Empirical Wavelet Transform,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: Moving into the specific details of "Tourist arrival forecasting with game-theoretic social influence weighting and deep learning," Jane, what are we looking at regarding the authors? Who came up with this work?

Jane: The paper is written by Ali Nikseresht from the School of Computer Science at Georgia Institute of Technology in Atlanta, and it's a collaboration.

Lu: Nikseresht is a senior researcher at Tsinghua University, and his team has been looking into representation learning for large language models. So you can tell he’s deep in the AI stuff.

Meng: He’s doing work on how to get these models to actually perform well when deployed, so I assume this paper is focused on making a model that works in the real world, not just a theoretical exercise.

Tom: Right. The abstract mentions it addresses forecasting international tourist arrivals in two specific contexts: Hong Kong from five major source nations pre-COVID-nineteen and Sanya in Hainan province post-COVID-nineteen <ref:2507.03411#pg1,Hong Kong from five major source nations pre-COVID-19, and>. That’s a big scope for one model to handle.

Jane: It really is; they are tackling two different types of tourism demand problems—pre and post pandemic scenarios—which makes the application quite diverse.

Lu: The paper highlights that they are using multiple sources of Internet big data, and specifically points out a gap in the literature where few papers have considered a mixture of multi-Internet big data in tourism demand forecasting twenty-six <ref:2507.03411#pg3,papers have considered a mixture of multi-Internet big data in tourism>.

Tom: That’s a key contribution, right? They are explicitly stating that using this mixture is the first major thing they contribute.

Jane: It suggests that just looking at volume or just valence isn't enough; you need to combine them.

Meng: I see why. If we only look at one signal, we’re probably missing half the picture of what a tourist is doing online.

The paper's summary: Tom: So, if we distill the abstract down to its simplest terms, what’s the actual problem they are solving in this paper?

Jane: They are trying to solve the problem of accurately forecasting how many international tourists will show up at a destination. It’s complicated because demand isn't simple; it has lots of non-linear dependencies, volatility, and seasonality that traditional models struggle with.

Lu: Traditional econometric models just can't keep up with all that complexity when you add in the noise from social media data and other big data sources.

Tom: And the solution they propose is a hybrid approach. They combine game theory for finding influencers, EWT to handle time patterns, and a stacked Bidirectional Long Short-Term Memory network for the final prediction.

Jane: So it’s about making the prediction more robust by layering these different techniques on top of each other to capture different parts of the demand signal that single models miss.

Meng: It sounds like a lot of components, which makes me wonder how much data they actually need to train all this up without running into massive computational issues.

Tom: They address that with the optimization part where they use Bayesian Optimization to tune the hyperparameters of their stacked BiLSTM network, which helps manage the complexity.

Jane: It’s a clever way to handle the nonstationary nature of tourism demand data, which keeps the model from becoming stale too quickly.

The paper's improvements: Tom: Let's focus on what they specifically claim are the improvements over what came before in this area. What makes this approach better?

Lu: Their first major contribution is utilizing a mixture of multi-Internet big data sources, considering both volume and valence features, which enhances decision-making efficiency.

Jane: And they also introduce the game theory based algorithm to identify opinion leaders using the Shapley value to calculate synergy. That’s a structured way to pinpoint influential communities instead of just looking at everything.

Tom: And then they layer in the Empirical Wavelet Transform, which handles cyclical, seasonal, and trend features in a way that gives them refined time-frequency analysis.

Meng: I’m interested in how this helps with deployment—does this reduce the computational load or just make it more complex to run?

Tom: The stacked Bidirectional Long Short-Term Memory network is designed to deal with nonstationary features, so the model itself is structured to handle those changing patterns better.

Jane: And they show that incorporating these elements together leads them to achieve lower error metrics compared to state-of-the-art models across various scenarios.

Conclusion: Tom: So, to wrap up the paper "Tourist arrival forecasting with game-theoretic social influence weighting and deep learning," what’s the final message they are trying to send?

Jane: The final message is that combining multi-Internet big data, game theory for opinion leader detection, EWT for time-frequency analysis, and a memory-aware stacked BiLSTM network creates a more accurate forecasting framework.

Lu: It suggests that this hybrid approach is effective when you look at the accuracy improvements across different statistical tests and scenarios for both Hong Kong and Sanya demand forecasts.

Meng: From an operational standpoint, it means businesses can use these kinds of data-driven methods to pinpoint sectors with the most significant impact on their specific demand fluctuations.

Jane: For policymakers, it suggests they should rely on these accurate predictions to plan infrastructure and resource allocation for places like Hong Kong and Sanya.

Tom: It’s about using this paper as a blueprint for building more sophisticated, data-driven systems in the tourism industry.

Georgia Institute of Technology

cs.LG, cs.GT, eess.SP

Submitted: 2025-07-04

Updated: 2026-10-14

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 60/100

The gist: The gist The proposed framework integrates game theory-based opinion leader detection, Empirical Wavelet Transform, and a Stacked BiLSTM network to generate accurate demand forecasts for

Key concepts

Game-Theoretic Opinion Leader Detection
This method uses game theory to find groups of social media users who have the maximum combined influence, or 'synergy.' It calculates a 'Trust Score' for users based on various trust metrics. The Shapley value is then used to precisely determine the contribution and synergy of each leader in a coalition, allowing the model to weigh influential opinions more heavily.
Empirical Wavelet Transform (EWT)
EWT is a mathematical tool used to analyze time-series data by breaking it down into different frequency components. This helps manage nonstationary attributes like trends and seasonality in tourist arrival data. By applying EWT, the model can effectively filter out cyclical patterns and extract meaningful features from the complex historical data.
Stacked Bi-directional Long Short-Term Memory (Stacked BiLSTM)
This is a deep learning architecture designed for sequence prediction. A standard LSTM processes information in one direction, while a Stacked BiLSTM uses two layers of LSTMs, processing data both forwards and backward. This structure allows the model to capture long-term dependencies and complex nonstationary features in the tourist arrival time series.
Bayesian Optimization (BO)
Bayesian Optimization is an intelligent search technique used to tune the hyperparameters of the Stacked BiLSTM network. Instead of randomly testing settings, BO intelligently chooses which hyperparameter combinations to test next based on previous results. This significantly reduces the time needed to find the best model configuration for accurate forecasting.

Terminology

Summary

The gist The proposed framework integrates game theory-based opinion leader detection, Empirical Wavelet Transform, and a Stacked BiLSTM network to generate accurate demand forecasts for international tourist arrivals in Hong Kong and Sanya

Data and Scope

The study addresses two distinct tourism demand cases: (i) arrivals to Hong Kong from five major source nations (pre-COVID-19), and (ii) arrivals to Sanya in Hainan province, China (post-COVID-19). The methodology utilizes multiple sources of Internet big data, including TripAdvisor, Google, Baidu, Facebook, and Twitter for the Hong Kong case. For the Sanya case specifically, only Twitter and Baidu are utilized as target Internet big data sources. The dataset for Sanya spans from January 2011 to October 2021, comprising a total of 130 data points.

Data Preprocessing and Feature Engineering

The data mining and wrangling framework involves two main phases: (i) data mining and wrangling and (ii) goal-oriented Internet big data analytics. This process includes steps such as defining the goals, processing the data, specifying required features, performing opinion leaders’ detection, assigning significance weights, performing Empirical Wavelet Transform (EWT), and injecting finalized data into the proposed hybrid forecasting approach. Preprocessed Internet big data for Hong Kong includes volume (attention) and valence (endorsement) features from Facebook, Twitter, and TripAdvisor. The normalization procedure for time series follows the formula given in Eq. (1).

Game-Theoretic Opinion Leader Detection

A novel game-theoretic approach is implemented to identify opinion leaders within social networks. This approach involves identifying a group of users having the maximum synergy, which is then designated as the coalition of opinion leaders. The algorithm utilizes a trustor-trustee tree generated on a basis of trust and centrality metrics to characterize user attributes. The calculation of the Trust Score (TS) involves computing Direct Trust (DT), In-Direct Trust (IDT), and Recommendation Trust (RT) metrics based on defined thresholds k, l, and s. The Shapley value is employed to calculate the synergy for all coalitions using Eq. (2A)

Hybrid Forecasting Model

The core of the prediction is a multi-staged algorithmic framework comprising game-theoretical detection, EWT, and the Stacked BiLSTM model. The process involves first detecting opinion leaders and assigning positive/negative weights to them based on their impact and significance. Second, EWT is executed on all data, including the Internet big data and tourism demand time series to compensate for cyclical, seasonal, and trend features. Finally, the ultimate forecasting is performed by implementing a novel algorithm which is the stacked version of the BiLSTM that deals with nonstationary features. The Stacked BiLSTM model optimizes its hyperparameters using Bayesian Optimization (BO) via the expected-improvementplus criterion to reduce search times and ensure superior forecasting performance.

Empirical Results and Contributions

The experimental results demonstrate that the proposed approach consistently outperforms existing algorithms, achieving the lowest error metrics among tested approaches. The study shows that incorporating Internet big data significantly improves forecasting accuracy and precision by a large margin. Specifically, integrating attention or endorsement features offers significant performance gains, with combining both yielding a remarkable uplift. The proposed hybrid algorithm proved superior to state-of-the-art models across various scenarios and statistical tests for both Hong Kong and Sanya demand forecasts. The theoretical contributions include pioneering the use of diverse Internet big data sources, alleviating seasonal, cyclical, trend dynamics through the memory-based algorithm and EWT usage, and utilizing game theory to identify opinion leaders.

Managerial Implications

From a managerial perspective, accurate tourism demand forecasting is critical for optimizing operations and improving profitability. Businesses can develop more accurate models by mixing various Internet big data sources, thereby demonstrating the true power and value of Internet big data in the tourism industry. The study suggests that incorporating opinion leader detection avoids a blanket examination of all social media posts by employing a game-theoretic approach to identify highly influential communities, aligning with the Pareto Principle. Accurate and timely forecasting results should be utilized by policymakers for planning destinations like Hong Kong and Sanya.

Future Work

Future research directions include exploring alternative optimization approaches for tuning hyperparameters to improve performance, expanding data collection to multilingual social media platforms, segmenting online communities using dynamic topic modeling or QAP, and adopting influential factor analysis methods such as SHAP and PDP for deeper interpretations of forecasting outputs. The study anticipates that these advancements will further underscore the utility of Internet big data within tourism management.

References

[1] C. Wu, X. Fu, and Y. Xia, “Spare part demand forecasting using PSO trained Quantile Regression Neural Network,” Comput. Ind. Eng., vol. 200, p. 110841, 2025

[2] N. Shahidian, P. Abreu, D. Santos, and A. Barbosa-Povoa, “Short-term forecasting of emergency medical services demand exploring machine learning,” Comput. Ind. Eng., vol. 200, p. 110765, 2025

[3] P. Rezaee and S. K. Moghaddam, “Assessment of design approaches for reconfigurable manufacturing systems based on forecasted demand data,” Comput. Ind. Eng., vol. 201, p. 110878, 2025

[4] A. Nikseresht, S. Shokouhyar, E. B. Tirkolaee, and N. Pishva, “Applications and emerging trends of blockchain technology in marketing to develop Industry 5.0 Businesses: A comprehensive survey and network analysis,” Internet of Things, vol. 28, p. 101401, 2024

[5] M. I. Sheshkol, K. Fardi, A. Hafezalkotob, R. Ogie, and S. Arisian, “Managing platelets supply chain under uncertainty: A two-stage collaborative robust programming approach,” Comput. Ind. Eng., vol. 198, p. 110645, 2024

[6] E. Doğruöz and R. Güllü, “Analysis of an all-or-nothing inventory model with price-dependent intermittent demand and supply uncertainty,” Comput. Ind. Eng., vol. 201, p. 110883, 2025

[7] Y. H. Liang, “Forecasting models for Taiwanese tourism demand after allowance for Mainland China tourists visiting Taiwan,” Comput. Ind. Eng., vol. 74, no. 1, 2014

[8] A. Nikseresht, D. Golmohammadi, and M. Zandieh, “Empirical modeling approaches in sustainable supply chain management: A review with bibliometric and network analyses,” Bus. Strateg. Environ., vol. 33, no. 8, pp. 8759–8783, Dec. 2024

[9] A. Nikseresht, D. Golmohammadi, and M. Zandieh, “Sustainable green logistics and remanufacturing: a bibliometric analysis and future research directions,” Int. J. Logist. Manag., vol. ahead-of-print, no. aheadof-print, Jan. 2023

[10] U. Gunter, I. Önder, and S. Gindl, “Exploring the predictive ability of LIKES of posts on the Facebook pages of four major city DMOs in Austria,” Tour. Econ., vol. 25, no.

Improvements for AI systems

  1. No preamble about improving AI systems is made; instead, focus on specific system enhancements based on the paper's contributions.

  2. The improved system can predict international tourist arrivals for Hong Kong (pre-COVID-19) and Sanya (post-COVID-19) by integrating a mixture of multi-Internet big data sources including TripAdvisor, Google, Baidu, Facebook, and Twitter.

  3. The system will identify influential user groups through a game theory based algorithm that assigns significance weights to opinion leaders based on synergy calculated using the Shapley value.

  4. The forecasting accuracy will be enhanced by incorporating Empirical Wavelet Transform (EWT) on all the Internet big data platforms and utilizing a memory-aware Stacked BiLSTM network.

  5. The system can manage nonstationary tourism demand attributes by applying EWT to capture latent sub-components like seasonality, cycles, trends, and an irregular term.

  6. The Stacked BiLSTM model, optimized via Bayesian Optimization (BO), will perform the final forecasting by leveraging memory features of our data and better dealing with the nonstationary features of the data.

  7. The system can discern how endorsement (valence) and attention (volume) signals influence predictive accuracy, allowing for a comparison between using all social media variables, only endorsement factors, or only attention elements.

  8. Policymakers in destinations like Hong Kong and Sanya can use the system to pinpoint the sectors with the most significant impact on destination demand fluctuations by identifying a coalition of opinion leaders.

Abstract

In the era of Industry 5.0, data-driven decision-making has become indispensable for planning in service systems such as tourism. This paper proposes a non-linear hybrid approach for forecasting tourist arrivals to Hong Kong from five major source markets before COVID-19 and to Sanya, China, through the COVID-19 shock. The method integrates multiple sources of Internet big data and employs a cooperative game-theoretic algorithm to identify coalitions of opinion leaders on social media, whose content then receives larger weights. The Empirical Wavelet Transform (EWT), applied causally, manages nonstationary attributes, and a Stacked Bi-directional Long Short-Term Memory (Stacked BiLSTM) network tuned by Bayesian optimization generates the forecasts. We prove that the Shapley values of the underlying trust-weighted coverage game have a closed form and that the coalition of maximal synergy can be found exactly in polynomial time. Relative to the same network fed only with past arrivals, Internet big data reduce the one-step-ahead root mean squared error by 4.1-41.7% across the Hong Kong markets and opinion-leader weighting by a further 10.1-49.4%, against 5.3% and 8.7% for Sanya. The approach compares favorably with published benchmarks and remains robust under volatile conditions, which suggests uses in logistics, supply chains and production planning.

Related papers