Along with the spread of social media, it has become possible to extract events occurring in the real world in real time. A benefit of analysis using data with position information is that it can accurately extract an event from a target area to be analyzed. However, because social media data include few data with location information, the amount to analyze is insufficient for almost all areas: we cannot fully extract most events. Therefore, efficient analytical methods must be devised for the accurate extraction of events with position information, even in areas with few data. For this study, we use geotagged tweets along with interpolation to estimate the best time to observe biological seasons when doing sightseeing such as cherry-blossom and autumn leaf viewing in areas and sightseeing spots. Herein, we explain the analysis results obtained using information interpolation and analysis of cherry blossoms in Japan during 2017.
INTRODUCTION
Because of the wide dissemination and rapid performance
improvement of various devices such as smart phones and
tablets, diverse and vast data are generated on the web.
Particularly, social networking services (SNSs) have
become popular because users can post data and various
messages easily. Twitter [
UISTDA '18, March 11, Tokyo, Japan. provide similar information in the form of guidebooks. Nevertheless, the information update frequency of that medium is low. Because each local government, tourism association, and travel company independently provides information about travel destination locales, it is difficult for tourists to collect information for “now” tourist spots. Therefore, providing current, useful, real-world information for travelers by capturing changes of information in accordance with the season and relevant time period of the tourism region is important for the travel industry. Tourist information for best times requires a peak period, which means that the best time is neither a period after or before falling flowers, but a precisely defined period to view blooming flowers. Furthermore, the best times differ among regions and locations. Therefore, for each region and location, it is necessary to estimate the best time for phenological observations. Estimating best-time viewing periods requires the collection of large amounts of information having real-time properties. For this study, we use Twitter data obtained from many users throughout Japan. We use Twitter, a typical microblogging service, and also use geotagged tweets that include position information sent in Japan to ascertain the best time (peak period) for biological season observation by region. The geotagged tweets are useful as social indicators reflecting real-world circumstances. They are a useful resource supporting a realtime regional tourist information system in the tourism field. Therefore, our proposed method might be an effective means of estimating the best time to view events other than biological seasonal observations.
To analyze information of each region from Twitter data, it is necessary to specify a location from tweet information. Geotagged tweets can identify places. Therefore, they are effective for analysis. However, because geotagged tweets account for only a very small proportion of the total information content of tweets, it is not possible to analyze all regions. We propose a method to provide tourists with information about sightseeing spots after processing a small amount of data in real time by interpolation using geotagged tweets.
RELATED WORKS
Along with rising SNS popularity, real-time information
has increased. Analysis using real time data has become
possible. Many studies have examined efficient methods for
analyzing large amounts of digital data. Some studies have
been conducted to predict real world phenomena using
large amounts of social big data. Phithakkitnukoon et al. [
In the following subsections, we describe the collection of
geotagged tweets to be analyzed, character preprocessing
for conducting analysis, and interpolation using Kriging.
Data collection
This section presents data collection. Geotagged tweets sent
from Twitter are a collection target. The range of geotagged
tweets includes the Japanese archipelago (120.0°E –
154.0°E, and 20.0°N – 47.0°N) as the collection target. The
collection of these data was done using a streaming API [
Next, we describe the number of collected data. According
to a report presented by Hashimoto et al. [
Preprocessing This section presents preprocessing. Preprocessing includes reverse geocoding and morphological analysis, as well as database storage for data collected through the processing described in the previous subsection.
From latitude and longitude information in the individually
collected tweets, reverse geocoding is useful to identify
prefectures and municipalities by town name. We use a
simple reverse geocoding service [
Morphological analysis divides the collected geotagged
tweet morphemes. We use the “Mecab” morphological
analyzer [
Preprocessing accomplishes necessary data storage for besttime viewing, as estimated based on results of the processing of the data collection, reverse geocoding, and morphological analysis. Data used for this study were the tweet ID, tweet post time, tweet text, morphological analysis result, latitude, and longitude.
Interpolation using Kriging
This section presents the method of interpolation, for which
we used Kriging [
We attempted estimation by interpolation using data aggregated for each spot. The estimated value of the target data at a point S0 is shown in formula (1) as a weighted average of the measured values Z(Si) (i = 1, 2..., N) at N points Si around point S0. Then we assigned value Z to tweets including the target word and Z. Here, N represents the 30 nearby targeted tweets. λ denotes a spherical model with decreased influence as distance increases. As described in this paper, weighting is done only by the number of tweets existing in the same spot. However, consideration of the weight of the tweet itself, such as using the number of retweets to tweets, is also necessary for future studies. :Unknown weighting of measured value at -th position 0 :Predicted position :Number of measurements EXPERIMENTS In this section, we explain the experiment for information interpolation for cherry blossoms in 2017, using the method described in the previous section.
We are conducting estimation experiments for cherry blossoms, autumn leaves, and other phenomena from 2015. As described herein, we used the period of cherry blossoms in 2017 while studying the interpolation method to improve estimation accuracy. The following subsections describe experimental datasets. Datasets Datasets used for this experiment were collected using streaming API, as described for data collection. The data, which include about 250
million items, are geotagged tweets from Japan during 2015/2/17 – 2017/5/13. The estimation experiment conducted to ascertain the best-time viewing of cherry blossoms uses the target word “cherry blossom,” which is “桜”, “さくら”, and “サクラ” in Japanese. We analyzed tweet texts that include the target word. About 100,000 tweets during the experiment period included the subject word.
The subject of the experiment was set as tourist spots in
Park,” “Shinjuku
Gyoen,” and “Rikugien.” Figure 1 presents the target area locations. A, B, C, and
D in the figure respectively denote “Takao
Memorial Park,” “Rikugien,” and “Shinjuku Gyoen.” In this experiment, about 30,000 tweets including the target word in Tokyo were found. In this experiment, all tweets made by the same user are also used as targets for analysis if they are tweets including the target word.
We conducted experiments of the following two kinds using these datasets. The first is an experiment using the number of tweets including the target word and the sightseeing spot name without interpolation.
experiment was compared as Baseline to confirm the usefulness of interpolation proposed in this paper. The second is an experiment using interpolation. In this experiment, we used Kriging in the earlier section. We present the example of Mt. Takao in Hachioji city. city from 2017/1/1 to 2017/5/13. The point denoted by A is Mt. Takao. Few geotagged tweets are related to cherry blossoms near Mt. Takao. For that reason, one cannot estimate the best time merely using tweets from A. Therefore, for this study, we interpolated the amount of information by Kriging using information of Hachioji city with Mt. Takao. Interpolation is done on a daily basis. The experiment results are presented in the next subsection.
A
B 16 km 32 km
C D 6 km
The light gray part of Figure 3 portrays an experimentally obtained result from interpolation results including the tourist spots. Apparently, A was able to produce an estimate using the proposed method by increasing the number of tweets using interpolation with surrounding tweets. For B and C, the useful information was included in the tweet not co-occurring with the tourist spot name. Therefore, we confirmed interpolation related to other kinds of cherry blossoms in early March of B and C and late April of C. In addition, for D, there are days when it can be determined more accurately by interpolating the number of tweets. Therefore, these results confirmed the possibility of estimating the peak period, even for an area without tweets, using data interpolation and overall tweet number interpolation.
CONCLUSION As described in this paper, we proposed an interpolation method to improve the accuracy of tourism information related to phenological observations. The results of the cherry blossom experiment conducted at the tourist spots in Tokyo in 2017 confirmed the trend of improved estimation accuracy using information interpolation. We confirmed the possibility of applying this proposed method to the estimation of viewpoints and sightseeing spots with few tweets. However, in regions with no geo-tagged tweets, another method must be considered. Research can be conducted in the future to verify whether similar results are obtained for other biological seasonal observations.
ACKNOWLEDGMENTS This work was supported by JSPS KAKENHI Grant Nos. 16K00157 and 16K16158, and by a Tokyo Metropolitan University Grant-in-Aid for Research on Priority Areas “Research on Social Big Data.”