| Abstract: |
Tourism studies, in particular those examining the distinctions between domestic and international visitors (DOM and INT), seldom use complex networks. This study suggests an approach based on transforming the time series of DOM and INT's itinerary total times (TM) and traveled distances (DST) into networks where the topology is quantitatively investigated. The Horizontal Visibility Graph (HVG) algorithm is used to turn these series into networks that retain some underlying system characteristics. The time series of yearly, monthly, weekly, and daily TM and DST are considered. These variables are calculated based on collected geotagged photos uploaded to the sharing platform Flickr and taken between 2008 and 2019 in the 23 special wards of the Tokyo Metropolitan Area.
The following research questions guide the study: (1) Relying on recorded decadal TM and DST, what are the characteristics of the resulting networks? (2) What such an approach can tell us about the major differences between DOM and INT? (3) Could TM and DST be used as proxies for representing and predicting their mobility trends? and (4) Do weather conditions influence such representativity and predictability?
We developed a four-step methodology to address these questions. First, we preprocessed the geotagged photographs to remove noisy records caused by malfunctioning hardware or GPS accuracy issues, then used machine learning to categorize users as local or foreign visitors. Second, using a convolutional neural network, photos were labeled into eight categories to consequently extract time- and distance-based clusters. This was followed by calculating and extracting time series representing the mean of yearly, monthly, weekly, and daily TM and DST recorded by both groups. The third step consisted of mapping the extracted time series into networks using HVG and conducting typical network analyses including mainly node degree distribution π(π) and the calculation of the scaling parameter π of π(π)=πβππ. Then, to investigate the resulted networksβ dynamics, we employed the assumption that suggests the critical value ππ=ππ(3/2) differentiate between stochastic (π>ππ) and chaotic (π<ππ) dynamics. The more stable and predictable the system is, the higher π of the exponential degree distribution. Moreover, for comparison and interpretation purposes, we employed four series with known dynamics as reference networks (random series, fractional Brownian motions, Logistic map, and the chaotic Lorenz map). Finally, we investigated possible weather influences by the previous process during bad and ideal conditions.
Preliminary results suggest that TM- and DST-based networks exhibit complex dynamics given the resulted exponential distributions. While all TM- and DST-based networks are ruled out for being chaotic systems with few exceptions, TM and DST are still close to the edge of chaos that could be similar to the behavior of random series. TM- and DST-based monthly routines of both DOM and INT during bad, ideal, and mixed weather conditions are much more predictable than those obtained daily and weekly. Overall, DST is found to be the suitable proxy for predicting INTβs daily routines, while TM is much more adequate for forecasting those of DOM. This study highlights the significance of combining social data and non-linear methods for further understanding touristsβ behaviors in general and the differences between those of DOM and INT in particular. |