PaperlayerView exact revision on arXiv

Source-linked AI summary

Using social network and semantic analysis to analyze online travel forums and forecast tourism demand

A Fronzetti Colladon, B Guardabascio, R Innarella

arXiv:2105.07727v1econ.EMcs.CLcs.SI

TL;DR

Tourism-demand forecasting matters to policy makers and tourism companies. This study analyzes TripAdvisor travel forums using social-network and semantic methods, finding that these variables could improve forecasts in 79% of cases and that models including them were in the superior set in 93%.

  • Problem

    The study asks whether online travel-forum activity can inform tourism-arrival forecasts, an issue relevant to policy makers and tourism companies.

    Source: §1 · pp. 3–4

  • Method

    The study analyzes TripAdvisor forum content with social-network and semantic methods, testing derived variables in existing forecasting models.

    Source: §1 · p. 3, §3 · p. 8

  • Results

    Social-network and semantic variables could improve forecasts in 79% of cases, and models including them were in the superior set in 93% of cases.

    Source: §5 · p. 23

  • Takeaways & Limitations

    Forum language complexity and communication centralization emerged as the most important predictors of international arrivals.

    Source: §5 · p. 23

    International airport arrivals are not a perfect proxy for tourism demand, though the authors consider them a reasonable indicator for this study.

    Source: §5 · p. 24

Abstract

from arXiv · show

Forecasting tourism demand has important implications for both policy makers and companies operating in the tourism industry. In this research, we applied methods and tools of social network and semantic analysis to study user-generated content retrieved from online communities which interacted on the TripAdvisor travel forum. We analyzed the forums of 7 major European capital cities, over a period of 10 years, collecting more than 2,660,000 posts, written by about 147,000 users. We present a new methodology of analysis of tourism-related big data and a set of variables which could be integrated into traditional forecasting models. We implemented Factor Augmented Autoregressive and Bridge models with social network and semantic variables which often led to a better forecasting performance than univariate models and models based on Google Trend data. Forum language complexity and the centralization of the communication network, i.e. the presence of eminent contributors, were the variables that contributed more to the forecasting of international airport arrivals.

1. Introduction

The study uses semantic and social-network analysis of TripAdvisor travel forums to develop variables for forecasting international airport arrivals in European capitals. It examines both post content and user interactions, while treating the analysis as exploratory rather than a full explanation of forecast influences.

1. Introduction

  • Online user-generated content can help anticipate shifts in tourist preferences and destination popularity through analysis of discussion topics and user interactions.The introduction also notes that people may rely on peers’ judgments and that sharing knowledge can make users feel part of a group and rewarded.Source: §1 · p. 3
  • The analysis covers 10 years of English-language TripAdvisor forum posts for seven major European capital cities, forecasting international airport arrivals.The authors suggest future work could use other online sources, languages, or tourism outcomes such as museum visits.Source: §1 · pp. 3–4
  • The study proposes content- and interaction-based variables from online communities that can be extracted from online sources and integrated into existing forecasting models.It investigates whether active communities with functional knowledge-sharing dynamics can help predict arrivals.Source: §1 · p. 3
  • The authors characterize the study as exploratory: it demonstrates informative value in semantic and social-network indicators but does not fully explain their influence on city-level forecasts.They call for dedicated future research to investigate the reasons behind those influences.Source: §1 · p. 4
  • More accurate tourism-demand forecasts can help policymakers and companies allocate resources, set pricing policies, and plan businesses while reducing misplanning risk.Source: §1 · p. 4
  • The proposed approach extends tourism forecasting beyond online reviews and web searches by applying social-network and semantic analysis to online-community data.The paper presents this as a new methodological approach and set of metrics.Source: §1 · p. 4

2. Forecasting Tourism Demand

The section reviews established tourism-demand forecasting approaches and motivates the use of online travel forums as a new source of social-network and language variables. It frames the study as addressing a gap in using forum dynamics to forecast international airport arrivals.

2.1. Exploring online community dynamics to predict tourism demand

  • Forums reveal specific information needs beyond common reviews and support analysis of European capital cities overall, while reviews may contain deceptive content.The authors choose forums rather than reviews to avoid limiting analysis to individual services or attractions.Source: §2 · p. 5, §2.1 · p. 6

2. Forecasting Tourism Demand

  • Tourism demand is often measured through arrivals and forecast using time series and seasonal trends; prior work also used search-engine and web-traffic data.Examples include a composite search index for Chinese tourism demand and autoregressive models combined with search-query data.Source: §2 · p. 5
  • Prior studies combined data sources and techniques to improve model accuracy, while the authors argue for new variables that integrate into existing models and are reasonably easy to extract in near real time.The passage cites approaches combining data mining with Markov chains and multiple online sources such as price levels and web traffic.Source: §2 · p. 6

2.1. Exploring online community dynamics to predict tourism demand

  • Few studies have used online travel forums or social-network analysis to forecast tourism demand; this study examines TripAdvisor-derived variables for forecasting international airport arrivals.The authors describe this as a gap and aim to assess social-network and semantic variables extracted from online big data.Source: §2.1 · p. 6
  • The study applies a three-dimension framework covering interaction, connectivity, and language, including network structure, change over time, response times, positivity, and complexity.These dimensions guide the analysis of social structure, interactivity, and language use in online conversations.Source: §2.1 · p. 7

3. Methodology

The study constructs monthly social, semantic, and search indicators from English-language TripAdvisor forum activity and user profiles, and evaluates their use in forecasting international airport arrivals across seven European capitals. It uses an adaptive pseudo-real-time setup to compare factor-augmented forecasts with autoregressive and Google Flight benchmarks, assessing accuracy across four horizons with mean square forecast error and model confidence set analysis.

3. Methodology

  • The analysis covered more than 2,660,000 posts by more than 147,000 users from January 2007 to December 2016, using English-language posts and participant profile data.The authors used English to keep semantic-variable measurement consistent and because it was the most-used language among tourists exchanging opinions across nationalities.Source: §3 · p. 8
  • Seven cities were selected for adequate forum data, and international airport arrivals—despite possible work-travel and transport-mode biases—served as a proxy for tourism demand.The sample comprised Amsterdam, Berlin, Lisbon, Madrid, Paris, Prague, and Vienna; Athens, London, and Rome were excluded because collection produced incomplete or inconsistent data.Source: §3 · pp. 9–10

3.1. Social Network Data

  • City forums were modeled as directed interaction networks, and the study measured connectivity, interactivity, and rotating leadership alongside language use.Measures included group degree and betweenness centrality, new users, activity, average response time, and changes in users’ betweenness centrality; few changes indicate static positions, such as eminent contributors, while more indicate shared leadership.Source: §3.1 · pp. 11–12
  • Language measures captured post sentiment and complexity, while Google Trends searches for city names paired with “flights” or “holidays” provided comparative predictors.Sentiment ranged from 0 (very negative) to 1 (very positive); complexity was based on word likelihood using TF-IDF information.Source: §3.1 · p. 13

3.2. Forecasting Model

  • To use the high-dimensional socio-semantic indicators, the authors augmented an autoregressive model with factors estimated by Partial Least Squares on the AR residuals.PLS incorporated information from both the target and predictors, while using residuals aimed to add information unexplained by the AR component.Source: §3.2 · pp. 14–15

3.2.1. Forecasting Procedure

  • Forecasts used monthly data from 2007:1–2016:12 in a rolling-window out-of-sample procedure, with lag and factor counts selected dynamically.Variables with unit roots were first-differenced, except Betweenness Group Centrality and Average Response Time; predictors were standardized for PLS.Source: §3.2.1 · pp. 15–16
  • The adaptive pseudo-real-time exercise evaluates forecasts from June 2013 to December 2016, with maximum autoregressive lag p=13 and at most 10 factors.Factor selection uses a 60-month window, with BIC or linear combinations explaining at least 20% of covariance between model (1) residuals and variables in Xt.Source: §3.2.1 · p. 16
  • Forecast comparisons include an adaptively lagged autoregressive benchmark, a Google Flight bridge model, and a factor-augmented bridge model combining Google Flight and SSI indicators.The FABM-GF factor is constructed from the errors of the BRIDGE-GF model.Source: §3.2.1 · p. 16
  • Forecast accuracy is compared using mean square forecast error at horizons h=1, 3, 6, and 12 months.The study identifies models forecasting equally well with the model confidence set analysis of Hansen, Lunde and Nason, testing equal predictive ability at 10% significance with a block bootstrap of 5,000 resamples.Source: §3.2.1 · pp. 16–17

4. Results

Social-network and semantic indicators improved forecasts of international airport arrivals, and language complexity and centrality were the most influential predictors. Results also show how forecast performance varied by horizon and city, and how forum users and discussions were characterized.

4. Results

  • Forum members were mostly men except in Paris, generally aged 35–64, and people posting tips and comments mostly lived in the cities discussed.The authors suggest younger users may seek information through other online sources, while noting that some may read forums without posting.Source: §4 · pp. 17–18
  • Discussions commonly concerned restaurants, hotels, attractions, and local transport; transport sentiment was negative, restaurant sentiment more positive than hotel sentiment, and museum sentiment positive except on prices and queues.Airport arrivals showed similar seasonality and often positive time trends across cities.Source: §4 · pp. 17–18
  • The preliminary Dumitrescu–Hurlin analysis found potentially significant arrival associations for all social-network and semantic metrics except average response time and group betweenness at lags 1 and 2.The test examined whether indicators could Granger-cause arrivals up to three months ahead after seasonal components were removed.Source: Table 3 · p. 18, §4 · pp. 18–19
  • Simpler substitutes for SSI indicators produced worse forecasts: SSI models outperformed them in 26 of 28 cases, and replacing activity and centrality metrics did not improve results.The simpler model was best only for Madrid at h=3, where it was included in the confidence set; adding the alternatives almost never improved BRIDGE-GF.Source: §4 · p. 20
  • Models with social-network and semantic indicators were best in 79% of cases and entered the superior set in 93% of cases.FAAR models without Google Flights outperformed other models in 54% of cases; six-month forecasts were weakest, with AR preferred for five cities, while indicator-augmented models did better for Amsterdam and Berlin.Source: §4 · p. 19, Table 4
  • Group betweenness, degree centrality, and language complexity most often ranked among FAAR’s most influential predictors across forecast horizons.New contacts gained importance for one-year forecasts, appearing in the IV quartile 54.76% of the time, while rotating leadership appeared there 38.10% of the time.Source: §4 · p. 22

5. Discussion, Limitations and Future Research

Forum-derived social-network and semantic variables improved tourism-arrival forecasts in many cases, with language complexity and communication centralization emerging as the strongest predictors. The discussion also qualifies airport arrivals as an imperfect but reasonable demand proxy and outlines future tests across methods, platforms, languages, and destinations.

Discussion, Limitations and Future Research

  • Male users predominated and most participants were 35–64 years old; requests about local transport were among the most recurring topics and had lower average sentiment.The authors attribute this in part to local-transport comments being commonly shared in forums rather than separate reviews.Source: §5 · p. 23
  • Social-network and semantic variables improved forecasts over Google Trend data in 79% of cases, and models including them were in the superior set in 93%; at six months, they did not improve AR models for 5 cities.Source: §5 · p. 23
  • Forum language complexity and communication centralization, measured by group degree and betweenness centrality, emerged as the most important predictors of international arrivals.Higher complexity may indicate new information and anticipate more arrivals; centralization may indicate eminent contributors or informal moderators who are local experts sharing knowledge with prospective tourists.Source: §5 · p. 23
  • Male-user share could improve forecasts; new-user counts were more informative than post volume, response time and user level were less important, and rotating leadership had high model weights in 50% of cases.Rotating leadership’s model weights declined at longer horizons; the authors suggest dominant local experts may support prospective tourists more than democratic interaction and plural opinions.Source: §5 · pp. 23–24
  • The authors propose testing these web-extracted variables with different forecasting techniques and investigating why their contributions differ.They argue that conventional tourism-demand models, including univariate arrival models, can be improved with online-source data.Source: §5 · p. 24
  • International airport arrivals omit visitors entering by other transport, work travelers, and travelers who continue elsewhere, but the authors consider arrivals a reasonable indicator for this study.Other candidate measures also have limitations, including visa counts that do not apply to EU citizens traveling within the EU and accommodation nights that are not always a good proxy for tourist numbers.Source: §5 · p. 24
  • Future research could test other social platforms, non-English sentiment and complexity, additional demand measures and visitor levels, cities beyond European capitals, and more resilient crawlers.The authors also propose studying major city-forum topics; Facebook groups may include more users under 35.Source: §5 · pp. 24–25

6. Conclusions

The conclusions argue that social-network and semantic analysis of online travel-forum data can improve forecasts of international tourist arrivals, extending tourism forecasting with forum-interaction metrics.

6. Conclusions

  • The forum metrics increased the forecasting accuracy of models that use online search-query volume when predicting international tourist arrivals.The authors conclude that these methods can help forecast international arrivals.Source: §6 · p. 25
  • The study contributes a new approach and metrics, addressing the limited exploration of travel-forum interactions compared with web searches and online reviews.The conclusions also describe tourism applications of social-network analysis as recent.Source: §6 · p. 25
  • Researchers, policy makers, and business managers can use forecasts for pricing and sales decisions; accurate forecasts also support planning and public investment.The conclusions cite avoiding overbookings or empty hotel rooms, appraising infrastructure projects, allocating resources, and setting medium-to-long-term tourism strategies as applications.Source: §6 · p. 25

Related papers

Loading 2105.07727v1…