1 Introduction

The change in the social interaction paradigm due to spread of social networks allows access to more data and additional information than traditional methods such as surveys, interviews or researches, as well as improves the interval time between sampling (Housley et al. 2014). Channels for expressing opinions have been increasing on a regular basis, and when these opinions are relevant to a company, they are important sources of business insight, whether they represent critical intelligence about customers, the impact of an influential reviewer on others purchase decisions, or early feedback on product releases, company news or competitors. Capturing and analyzing these opinions is a necessity for proactive product planning, marketing and customer service, and it is also critical in maintaining brand integrity. The importance of harnessing opinion is growing as consumers use technologies such as Twitter to express their views directly to other consumers (Kalarikkal and Remya 2015). Recently, several studies have been developed that aim to characterize the communities formed in social networks by analyzing the use of certain symbols or own conventions of writing small messages on Twitter to disseminate information, to express emotions or to make a topic a major issue in social network. The study presented by Honeycutt and Herring (2009) focused on the importance of using the @ symbol in British tweets as a form of direct interaction with other users. Hong et al. (2011) compared the behavior and use of Twitter by various communities identified in a corpus made up of 62 M tweets. This work identified the top 10 most used languages on Twitter, extended Honeycutt’s work (Honeycutt and Herring 2009), and looked for patterns of sharing URLs, hashtags, mentions, replies and retweets. Boyd et al. (2010) examined the dissemination of information on Twitter by retweeting action, and Burnap et al. (2014) presented some models to predict information flow size and survival using data derived from Twitter by defining information flows as the propagation over time of information posted to Twitter via the retweeting action. Bora et al. (2015) analyzed the ability to predict the emergence of virality of a hashtag. Java et al. (2006) identified several categories of intention to use Twitter including: (1) daily chatter where users discuss events in their lives or their current thoughts; (2) sharing information or URLs and (3) reporting news which includes commenting on current events or automated news agents posting weather or news stories.

In addition to the knowledge and interpretation of behavior in social networks, many such applications could benefit from information about the location of users, but only less than 1 % of tweets are geotagged and information available from the location fields in users’ profiles is scarce. Mahmud et al. (2014) present some algorithms to predict the home location at the city level of Twitter users from the content of their tweets and their tweeting behavior. They also examined the possibility of predicting at other larger levels of granularity, such as state, time zone and geographic region. Other authors built geographic topic models to predict the location of Twitter users in terms of regions and states (Eisenstein et al. 2010). Hecht et al. (2011) built Bayesian probabilistic models from words in tweets for estimating the country and state-level location of Twitter users. They used location information submitted by users in their Twitter profiles, resolved via the Google Geolocation API, to form the ground-truth of a statistical model for location estimation. Cheng et al. (2010) describe a city-level location estimation algorithm, which is based on identifying local words from tweets and building statistical predictive models from them, but their method requires a manual selection of such local words for training a supervised classification model. Chandra et al. (2011) described location estimation using the conversation relationship of Twitter users in addition to the text content used in the conversation. Chang and Sun (2011) described yet another content-based location detection method using Gaussian Mixture Model (GMM) and the Maximum Likelihood Estimation (MLE). Their method also eliminates noisy data from tweet content using the notion of nonlocalness and geometric localness.

Capturing human movement patterns across political borders is difficult. Blanford et al. (2015) analyzed 10 months of geolocated tweets for Kenya and studied movement of people at different temporal (daily to periodic) and spatial (local, national to international) scales. The use of geolocated tweets is also reported in Widener and Li (2014) in order to present a study about the consumption of healthy and unhealthy foods by the US population. Tweets with known location are also used by Saravanan et al. (2013) and Kim et al. (2013) for real-time information on the most relevant topics covered by users, by conducting a review of feelings indicating if a discussed topic is positive or negative. A methodology by which it is possible to discover the occurrence of a relevant event in a certain place, by collecting and analyzing geolocated tweets is proposed by Kim et al. (2011). Another recent and interesting study uses 2 years of geolocated data from Twitter to track trends in migration patterns (Zagheni et al. 2014). The paper shows that publicly available geolocated tweets can, without additional information, help to understand the relationships between internal and external migration. Other related work includes a method presented by Culotta et al. (2015) to predict the particular user location, based on the user’s followers. An analysis on how geolocated information coming from cellular data can help monitoring and mapping spatial and temporal variability of population in a specific region can also be found in Manfredini et al. (2011).

Most of the above-referred studies base their analysis on tweets published in English (Hong et al. 2011). To the best of our knowledge, our work is the first to systematically study how Portuguese users behave in Twitter, more specifically in each of the Portuguese districts. This work uses a database of geolocated tweets, produced in Portugal during a 1-year period, and written in European Portuguese. The database was created using several strategies for overcoming some of the Twitter API limits (Brogueira et al. 2015, 2016). We use information about a tweet’s date and time to analyze the distinct Portuguese regions in terms of the number of tweets produced at a given period of time (day, season, etc.). Such data can be used to provide insights about the “connected” Portuguese population, such as adoption of new technologies, population distribution, main interests, or mobility. We also analyze tweet contents in order to attempt to characterize Twitter usage within the Portuguese community and compare it with other communities.

Our research is motivated by the attempt to characterize the use of Twitter in Portugal and results in the following main contributions:

  • An algorithm for predicting the Portuguese districts where each tweet was published;

  • Pattern identification about Twitter usage during work periods and holiday periods;

  • Characterization of Twitter usage by the Portuguese community (based on hashtag, mention, retweet, replies and URL use);

  • Characterization of district sentiment based on the analysis of the frequency of emoticons.

The remainder of this paper is organized as follows: Sect. 2 describes Twitter’s conventions. Section 3 presents the methodology for data acquisition and processing. Section 4 presents the temporal and geographical analysis of collected data. Section 5 addresses topics and results concerning the characterization of geolocated Twitter usage in Portugal. Section 6 presents a small case study on Portuguese tourists in the Algarve region that shows an example of applicability. Finally, Sect. 6 presents the major conclusions and prospects for future work.

2 Background

2.1 Twitter

Twitter is a microblogging service based on short messages limited to 140 characters called tweets. Twitter has currently around 320 million active usersFootnote 1 that publish about 500 million messages per day.Footnote 2 The freedom to share thoughts and opinions about different aspects of daily life, feelings or news about various subjects, makes the volume of information present on Twitter, potentially interesting for several studies in diverse areas such as policy (Rill et al. 2014), tourism, marketing or health (Culotta 2014; Santos and Matos 2013). Twitter data have also fueled the rapid growth of consumer-generated content such as consumer satisfaction, opinion extraction, ratings and sentiment analysis (Kalarikkal and Remya 2015). Furthermore, research suggests that the online purchase intent is significantly impacted by negative/positive sentiments found online, and Twitter data as been used, for example, to understand public mood and use the predicted mood values to infer the stock market movements (Mittal and Goel 2011).

An hinder to perform such analysis is data access. Despite Twitter being a public social network, the Twitter API provides only a limited access to the total volume of produced tweets. For example, the Streaming API accesses in real time a continuous stream of tweets that, depending on the level of used permission authentication (Kumar et al. 2014), corresponds to 1 % of the total tweets produced at a given time. Alternative APIs limit the access in other ways.

2.2 Categories of users and intention to use Twitter

Twitter is used for purposes as diverse as: (1) sharing ideas and thoughts; (2) information dissemination and news; and (3) communication or conversation with friends. Java et al. (2006) identified three main categories of Twitter users: information sources, friends and information seekers. Information sources post news and tend to have a large base of “followers.” These sources may be individuals or automated services. Friends are a broad category that encompasses most users, including family, co-workers and strangers. Information seekers tend to be users who may post rarely but who follow others regularly. As previously mentioned, Java et al. (2006) also identified several categories of intention to use Twitter, including: (1) daily chatter, where users discuss events in their lives or their current thoughts; (2) sharing information or URLs; and (3) reporting news, which includes commenting on current events or automated news agents posting weather or news stories. Other category of user intention is conversation (Java et al. 2006), a quite popular use within the Twitter Portuguese community, as we reveal later in this article.

2.3 Twitter conventions

Each category of users or type of Twitter usage is characterized by the use of certain Twitter symbols. Twitter users can refer to a specific user by including a mention anywhere in their tweets, done in the form of @username, where username is the mentioned user’s screen_name. The information about all mentions contained in a single tweet is presented in the field entities.mentions. A reply to a previous tweet is a specific form of mention, with the @username appearing at the beginning of the reply tweet. Retweeting is typically used to spread information received from friends to followers (Boyd et al. 2010). It is the equivalent of forwarding an email and is an action to information sharing and a social act, recognizing and promoting someone’s message. A common form of retweeting is “RT @username message,” where “message” is a tweet created by “@username.” Users have also adopted a variety of other syntactical markers such as “RT:@,” “retweeting @,” “retweet @,” “(via @),” “RT (via @),” “thx @,” “HT @” and “r @” (Boyd et al. 2010).

Hashtags are keywords included in tweets, in the form of #keyword. Including a hashtag creates a tag for a social bookmarking system and specifies a mechanism useful to collect tweets related to the given topic suggested by the keyword. The field entities.hashtags contains the information about all hashtags present within the message.

In order to share information, Twitter users can include URLs or links in their tweets. The information about any used URLs is presented in the field entities.urls.

3 Methodology

3.1 Data acquisition and processing

The dataset used in the scope of this paper was collected between September 15, 2014, and September 14, 2015, using the Streaming API filter/status and considering only the collection of tweets produced in Portugal and written in European Portuguese. The data were collected by a Python script that directly accesses the Twitter API and was restricted to the geographic limits corresponding to the Portuguese mainland and the Autonomous Regions of Azores and Madeira. Additionally, the tweets were also restricted to those in which the language field lang, automatically assigned by Twitter, contains the value ‘pt’ and the place.country field contains the value “Portugal.”

The Streaming API allows the collection of up to 1 % of all published tweets. Using the Streaming API results to perform data analysis can raise issues concerning the validity of the data due to the quality of the sample and any eventual bias. However, by combining the Streaming API with filters to capture geolocated tweets produced on a delimited geographical area, it is possible not only to capture most of the geolocated tweets produced in that area and easily circumvent the 1 % limit, but also obtain an unbiased sample of the overall twitter production within that area in what concerns, for example, the discussed topics, as shown by Malik et al. (2015) and Morstatter et al. (2013). As such, we believe that the data we collected are unbiased and can be used to extract indicators regarding the overall Twitter use in Portugal, and for some demographic indicators (e.g., population per district) for the general population, even if such generalization should not be assumed a priori.

The final collection contains about 27.8 M tweets produced by 97,584 users and is stored in a large MongoDB database.

The information about each published tweet contains not only the message, but also author information and location at the time of the post. One of the goals of this work concerns the analysis of each of the 20 administrative districts within Portugal mainland. Therefore, all results depend on how well we can assign the location where a given tweet was produced to the corresponding district. However, most of the times such information cannot be easily retrieved from the tweet. The remainder of this section describes the approach used in assigning the district to the location where the tweet was produced.

3.2 District inference by locality name

The information contained in each tweet has a flexible schema, and the data about the author and the location where it was produced are included in documents imbedded in the tweet structure. All geographically localized tweets contain the embedded document “place,” which assembles a number of fields that provides, as a whole, information about the geographical location where the tweet was produced. Such information can be found in the place.name and place.full_name fields. In some cases, place.full_name contains not only the location, but also the country to which the location belongs. For instance, with the value of the field “Lisboa, Portugal”, the reference to “Lisboa” is the name of the city Lisbon and the reference to “Portugal” is the country name to which Lisbon belongs. Using the information found on the place.full_name field, we developed a method to obtain the district name based on the locality name (Fig. 1).

Fig. 1
figure 1

Algorithm to infer the district of each locality

Full size image

This method involves several steps and is based on the list of postal codesFootnote 3 provided by CTT—Correios de Portugal SA (Portugal’s postal service). The list contains, among other information, the association between the locality and the district for all locations of the Portuguese mainland, Azores and Madeira. The information is stored as CSV (comma separated value) files, where each line contains 16 data fields separated by semicolons, including the following information: district code; county code; locality code; and locality name. Table 1 shows an example of such an entry, where “01” corresponds to Aveiro district, “04” is the code of Arouca municipality, and 69,893 is the code of Picoto, the corresponding location. The district and municipality codes are also available as separated files.

Table 1 Match of district and municipality codes
Full size table

The process begins by checking whether the name of the settlement indicated in the first value of the place.full_name field is included in the list of locations of CTT—Correios de Portugal SA. If not, we consider the possibility that the location name contains a typographical error, and we use a dictionary containing the most frequent errors in order find the probable matching district. For example, if the location shows the values of “Lisbon” or “lisbonne,” we consider that the respective tweets are deemed to have been published in the town of Lisbon (and as such, in the district of Lisbon). Location name typographical errors occurred in around 2.93 % of the database tweets. When the location name is correctly written and is present in the list of locations of CTT—Correios de Portugal SA, finding the corresponding district is a direct task unless the name occurs in different districts. However, 47 % of the tweets refer locations names that exist in more than a single district. One example is “Covilhã,” a well-known (in Portugal) city in the Castelo Branco district, but whose name is also associated with two smaller localities in Porto and Braga districts. Another example is the city “Seixal,” which is both a city in Setúbal district, and a much smaller locality in Aveiro district. These examples are shown in Table 2.

Table 2 Examples of localities sharing the same name
Full size table

In 43 % of the tweets, the relative size of the locations is substantially different, and in such cases we assume that the tweet is originated from the locality with the much larger area. To calculate the area of each place, it was considered the number of matching streets, being considered that a particular locality has a much greater area if has a much larger number of streets listed in the database. If the location exists in more than 2 districts, one locality must be much larger than all others (Table 3).

Table 3 Examples of localities sharing the same name
Full size table

When the locations of the two largest values are not too dissimilar, we solve the ambiguity by taking into account the geographical coordinates present in the field geo.coordinates (if different from null). This is accomplished by invoking a service of the Google Maps API that given a pair of coordinates (latitude, longitude) lets you know which town and the district correspond to the coordinate pair. Also for cases where the place.full_name field contains the value of “Portugal” invoking the Google Maps API service allows you to get the name of the district, which otherwise would be impossible to determine. Table 4 shows an example of this invocation and the respective Google Maps API return.

Table 4 Using Google maps API
Full size table

The reason why the disambiguation process starts by comparing locality area size instead of using the geographical coordinates lies in the fact that the Google Maps API imposes a maximum limit to the number of daily invocations from a given IP address. Due to the size of the used corpus, it would be impossible to use the API to determine the district for all ambiguous cases. Using our method, the Google Maps API is only invoked in 0.03 % of the tweets (corresponding to the disambiguation of cases where the place.full_name field is not mentioned, or the tweet was posted in a location that can belong to different districts and have a similar areas).

The previous steps allowed for resolving the district in 99.32 % of tweets. In the remaining 0.68 % of the corpus, the place.full_name field did not contain any locality name and instead of precise coordinates it is presented a geographic. In such cases, we used the center of the referred area and the Google Maps API to find the corresponding district.

4 Temporal and geographical data analysis

The Portuguese mainland is divided into 18 districts: Aveiro, Beja, Braga, Bragança, Castelo Branco, Coimbra, Évora, Faro, Guarda, Leiria, Lisboa, Portalegre, Porto, Santarém, Setúbal, Viana do Castelo, Vila Real, Viseu. The Madeira archipelago is composed by Madeira and Porto Santo islands. The Azores archipelago consists of nine islands: Santa Maria, São Miguel, Terceira, Graciosa, São Jorge, Pico, Faial, Corvo e Flores. For representation clarity and due to the population size, we grouped the islands into two districts corresponding to each archipelago: Madeira and Azores.

The Portuguese population is not equally distributed in the Portuguese territory (Instituto Nacional de Estatística 2011). A sharp desertification is noticed in large areas of interior, and a high population density can be found on the coast and metropolitan areas, in particular Lisbon and Oporto. Census 2011 (Instituto Nacional de Estatística 2011) also refers to the distribution of young and elderly population: The coastline contains a superior percentage of young people. The situation is reversed in relation to the elderly population. As such, it is common to divide the territory into coastal and interior regions when performing data analysis (Fig. 3a).

Figure 2 shows the Portuguese population per district and the distribution of Twitter users together with their activity for the Portuguese territory, based on our database of tweets. It is clear a high correlation between population and Twitter use. Most users are active in the coastal districts of Portugal, particularly in Lisboa (~44 K), Porto (~23 K), Setúbal (~15 K) and Faro (~12 K). Faro district corresponds to the Algarve region, and as such, its number of users is inflated by influx of population during the summer holiday period. Map c shows that the most active users are in the districts of Aveiro, Setúbal, and Lisboa.

Fig. 2
figure 2

Distribution per district of the Portuguese population, the Twitter users, and also the average number of tweets per user per day. a Population in Census 2011 (https://pt.wikipedia.org/wiki/XV_Recenseamento_Geral_da_População_de_Portugal, last accessed November 15, 2015). b Twitter users per district. c Average tweets per user

Full size image

The Portuguese Twitter community is essentially composed of teenagers or young adults (Brogueira et al. 2014a, b), which given the highest percentage of young people on the coast of Portugal, partly explains the higher volume of tweets collected in coastal districts as well as the higher user activity. The largest tweet production occurs in Lisboa (~8.8 M), followed by Porto (~4.3 M) and Setúbal (~3.2 M), which is consistent with the distribution of the Portuguese (Instituto Nacional de Estatística 2011).

In addition to the coast/interior differences, it is also usual to look for regional differences between the North, the Center and the South. The North region includes the districts of Aveiro, Braga, Bragança, Guarda, Porto, Viana do Castelo, Vila Real, and Viseu; the Center region contains the districts of Castelo Branco, Coimbra, Leiria, Lisbon, Portalegre and Santarém; and the South contains districts of Beja, Évora, Faro, and Setúbal (Fig. 3b).

Fig. 3
figure 3

Common divisions used for analysis of Portuguese regional data. a Coast (dark) and interior (light) districts. b North, Center and South districts

Full size image

A significant part of the population usually moves away from the major urban areas to their homelands or to the leisure areas (mostly to Algarve). In order to account for any seasonal regional population flows, we also decided to divide our analysis into eight working and holiday periods according to the 2014/2015 school calendar as defined by the Ministry of Education and Science of the Portugal Government.Footnote 4 The working periods consisted of: September 15 to December 16, 2014; January 6 to February 15, 2015; February 19 to March 20, 2015; and April 7 to June 12, 2015. The Christmas holiday period and New Year’s Eve was considered from December 17, 2014 to January 5, 2015; the Carnival holiday period from February 16 to 18, 2015; the Easter holiday period from March 21 to April 6, 2015; and the Summer holiday period was considered from June 13 to September 14, 2015.

An analysis of daily Twitter usage between the different areas (Coast/Center; North/Center/South) does not show any significant differences especially during working periods. However, the activity profile is significantly different between work periods and holidays. Figures 4 and 5 show, respectively, the Twitter activity throughout the day during work and holiday periods for Coast/Center and North/Center/South.

Fig. 4
figure 4

Daily activity in interior regions and coastal regions. a During a work period. b During holidays

Full size image
Fig. 5
figure 5

Daily activity for the North, Center and South. a During a work period. b During holidays

Full size image

During the work periods, Twitter activity distribution is quite similar throughout the country, without significant regional differences. The activity reaches a minimum between 3:00 and 5:00, has a constant growth rate between 7:00 and 12:00 (lunch break start), has a slight decay during and after lunch hours, and grows at a maximum rate between 16:00 and 21:00, when it reaches peak hour usage. Activity then decreases more or less constantly until the mentioned minimum activity hour.

Daily activity during holiday periods is quite different from the work periods, even if it is still possible to observe the lunch hour peak and the minimum and maximum activities occur roughly around the same time of day. One of the most notorious differences is the extended activity during the night period. Instead of peaking around 21:00–22:00 and rapidly decreasing, holiday activity extends throughout the night: activity until 2:00 in the morning is higher than lunch time activity; and the minimum activity period is between 6:00 and 7:00, i.e., 1–2 h later than during working periods. The lunch time peak also occurs 1 h later than during the work periods.

Regional differences are also more significant during holiday periods than during work periods, which indicate that, as expected, certain areas are more vacation oriented than others. Figure 6 summarizes the activity in all working and holiday periods per district (data are not normalized by period duration). Figure 7 shows the same information but aggregating all work periods and normalizing each period by its duration. The first conclusion to obtain from these charts is that twitter activity is lower during work periods than during holiday periods and that the most prolific season for tweet production in Portugal is Carnival, a period of days largely associated with partying and celebration.

Fig. 6
figure 6

Twitter activity during work and holiday periods

Full size image
Fig. 7
figure 7

Average number of tweets per day (as a percentage) for each considered period. Twitter production is lowest during work periods and reaches its highest during Carnival

Full size image

Regional differences are visible during the summer period. Regions that are common vacation destinations (such as Algarve, district of Faro) exhibit an expected activity increase. But the most interesting seasonal differences are seen in regions that receive a large influx of emigrants visiting their families (Guarda, Viana do Castelo, Azores). Such regions have an aging (and nontechnological) population that is transformed with the seasonal arriving of the emigrants and their child. This results in a very noticeable Twitter activity increase. Note that the summer period lasts from mid-June to mid-September, but most vacationing emigrants stay only for a 1-month period. Their impact would be even more visible in that period.

The information presented in the seasonal charts can prove to be useful when deciding the timing of launching advertising campaigns, or for a more efficient dissemination of news targeted to the interests of the population in each region. The same is valid in what concerns the knowledge of the time of the day when the target audience in a certain region is more active on Twitter: The information can be propagated more effectively and viewed by a larger number of potential customers.

5 Data characterization

5.1 Usage of Twitter conventions

The proportion of occurrence of Twitter conventions, i.e., URLs, hashtags, mentions, replies and retweets, allows the characterization of Twitter use on a particular community. A study by Hong et al. (2011) performed such analysis for different languages. Among others, it included the study of 6 M tweets in Portuguese language (mostly Brazilian Portuguese), and 31 M English tweets collected between April 18 and May 16, 2010. The first lines in Table 5 present the study results, and show a high prevalence of conventions. For example, in the case of the Portuguese language tweets, the study found that 13 % of tweets contained URLs, 12 % contained hashtags, 50 % of tweets made reference to other users via mentions, 32 % of tweets were replies to other tweets, and 12 % were retweets. It should be emphasized that the tweets used in the study did not consider geolocalization and were not related to specific countries, but to specific languages.

Table 5 Percentage of tweets using various conventions (English and Portuguese languages, and geolocated in Portugal and in Portuguese)
Full size table

A similar analysis was performed on our geolocated corpus of 27.8 M Portuguese tweets. Table 5 shows the results we obtained for the whole country and per district. The results we found are noticeably different, showing a much lower usage for each category (in some cases one order of magnitude less!). These large differences are surprising but can be explained (or at least attenuated) by some factors, the most important of which is the fact that all the tweets in our database are geolocalized. This is a very relevant issue because most Portuguese mass tweet producers, such as news agencies, newspapers, TV channels or TV shows, do not publish geolocalized tweets, and as such are absent from our database. The only notorious exception is the newspaper “Jornal de Notícias” (Table 6). Tweets by such users are characterized by including conventions: They incentivize the use of #hashtags, often include URLs to sites where the readers can find more details on the message they are commenting or transmitting, and often retweet related messages (with or without recurring to mentions). The weight of the mass tweet producers within all the produced tweets can explain the much lower Twitter convention values we found, but also gives the important indicator that the Portuguese community using geolocation is not adept at using Twitter conventions especially in what concerns information dissemination. Less important reasons for the differences include the fact that in Hong et al. (2011) most of the tweets are from Brazilian users (which at the time outnumbered the Portuguese Twitter community by a large factor: 30 M users vs. less than 600 K), and Twitter Brazilian and Portuguese communities do not necessarily behave similarly.

Table 6 Jornal de Notícias, the only major Portuguese newspaper that tweets using geolocalized information
Full size table

Analyzing the data presented in Table 5 while taking into consideration the categories of Twitter use suggested by Java et al. (2006), it is possible to state that the Portuguese community on Twitter uses this social network essentially to chat, exchange thoughts and share opinions. This is supported by the fact that the conventions with a significant usage percentage are mentions (17.88 %) and replies (16.97 %), while the conventions associated with the dissemination of information, URLs (2.13 %) and retweets (0.02 %) have a residual usage.

Differences between districts exist but are not very significant: Mentions and replies are less used in the interior (which might or not indicate a lesser tendency for conversational use in the interior); URLs are strangely more frequently in Viana do Castelo; Lisbon users seem to be more adept at hashtagging than habitants from other regions, etc.

5.1.1 Topics of conversation in Portugal: hashtags

In terms of the content of messages posted in each of the districts during the period of September 2014 to September 2015, it was found that the most used hashtags in each district involve football (soccer) and entertainment prime time TV shows. The most used #hashtag is common to all districts, #carregabenfica, which is related to “Sport Lisboa e Benfica,” the Portuguese football champion in 2014/2015. The hashtags #ss5, #idolospt and #unicamulher refer to prime time entertainment TV programs and are also top hashtags in all districts. Figure 8 shows the frequencies of the top 4 hashtags common to all districts. The discussion and exchange of views on football and television programs cut across all districts of Portugal, but the theme of football has prevalence, since among the top-k hashtags for each district are #diadesporting, #sportingcp, #somosporto, #fcporto, #rumoao34, all related to the three largest Portuguese football clubs.

Fig. 8
figure 8

Occurrence of the top 4 hashtags in all districts

Full size image

5.1.2 Information share: URLs

Twitter users can include URLs or links in their tweets with intention to share, or elaborate, on the information published on a tweet. The top 5 domains shared in the analyzed corpus of tweets are as follows: 1—www.instagram.com; 2—www.trendinalia.com; 3—www.dlvr.it; 4—www.swarmapp.com; 5—www.youtube.com.

Compared to the tweets in English analyzed in Hong et al. (2011), the only top common link is the sharing of videos from YouTube, even though Instagram also appears in many common lists. No significant regional differences were observed in the top 5 positions, but going lower it is possible to find differences, such as, for example, a local car dealership in the district of Braga that uses Twitter to promote its products.

Sharing URLs is directly related to information sharing. Figure 9 analyzes the frequency of occurrence of URLs related with each of the main Portuguese newspapers. As mentioned before, “Jornal de Notícias” is the only major newspaper that publishes on Twitter using geolocalized information (Table 6). As such, it is no wonder that “Jornal de Notícias” is by far the most frequent URL in our database (Fig. 9a). Note that the “Correio da Manhã” newspaper’s website got in July 2015 about five times more visits than the site “Jornal de Notícias,Footnote 5 and as such it should be much more popular in Twitter if one excludes the geolocation bias. The same source states that in terms of sports newspapers, the website of newspaper “aBola” is the leader in both visits and previews, followed by the websites from “Record” and “O Jogo.” This trend is confirmed in Fig. 9b. In this case, there is no bias since none of the three sports newspapers publishes using geolocalized information.

Fig. 9
figure 9

Newspapers URLs shared by geolocated Twitter users. a Generalist newspapers. b Sports newspapers

Full size image

5.2 Emoticons

Individual happiness is a fundamental societal metric (Dodds et al. 2011), and as such one factor worth of analysis using any available data. Hence we decided to perform some exploratory sentiment analysis. The most commonly used techniques for sentiment analysis are sentiment lexicons. However, these resources are still scarce for the Portuguese language. Moreover, some of the existing lexicons for Portuguese correspond to automatic translations (by Google translate) from their English originals. This usually leads to low-quality sentiment analysis results in Portuguese texts. It should also be noted that the applicability of sentiment lexicons to tweets can be limited because the existing lexicons are not built for the Twitter domain and therefore do not tackle well typos and other specific phenomena commonly found in tweets. Since emoticons are naturally expressed in tweets, and more than 15 % of tweets in the used corpus contain emoticons, we opted to use them in this task.

Emoticons allow expressing a large and diverse set of emotions using a compact representation that uses just a couple of characters, and the Twitter 140 character message size limit incentives a their use. Therefore, emoticons are a priori a means to characterize user happiness in Twitter.

We looked for emoticons within our geolocalized database and categorized them into “positive,” “neutral” and “negative” feelings. Table 7 presents the top 10 emoticons for each category and respective occurrence frequency in the database. The results were rather surprising: The most frequent emoticons (by a large margin), “:/” and “0/”, are used to express “confusion” and “frustration,” The occurrence of staples such as “:)” or “:(” is one order of magnitude lower than “:/” or “0/”. This can be seen as a huge indicator of the young age of most Twitter users in Portugal and shows how they are lost about their future and their lack of perspectives under the crisis and austerity affecting the country during the analyzed period. The results also reveal the lack of anger and revolt usually associated with youngsters that are more politically oriented and use Twitter for dissemination of ideas, reinforcing the notion that Twitter is mostly used in Portugal for more soft and/or recreational purposes.

Table 7 Top 10 emoticons used for representing positive, neutral and negative feelings and respective frequency
Full size table

Figure 10 shows how a positive, b neutral and c negative districts are. The darker the tone, the more intense is the feeling. Intensities are comparable between the figures and show the prevalence of negative feelings and a generalized degree of frustration, dissatisfaction and unhappiness (tone intensity is higher for negative feelings across all districts), which, once again is in line with the overall sentiment associated with the austerity imposed to the country during the period in question.

Fig. 10
figure 10

Emoticon frequency per district. a Positive feelings. b Neutral feelings. c Negative feelings

Full size image

5.3 Tweet publishing source

Each tweet contains a “source” field that is used to indicate the type of device used for the tweet publication. In addition it is possible to know the device operating system (e.g., Android, Windows Phone, iPhone). An analysis by district shows that Android is the most prevalent source for geolocalized tweet publishing in all Portuguese districts. Figure 11 shows the top publishing sources used in each district.

Fig. 11
figure 11

Top Twitter publishing sources per district

Full size image

A possible interesting analysis on this data arises from Edwards (2014) statement in Business Insider: “the rich, it seems, use iPhones while the poor tweet from Androids.” If one accepts such statement as true—which is obviously very debatable—it would be possible to infer the overall wealth level of the country (more “poor” than “rich”), and an analysis based on iPhone usage for tweeting could indicate which are the “wealthiest” Portuguese regions. As can be seen in Fig. 12, this statement does not seem to hold in the case of Portugal, since districts such as Bragança, and Vila Real, which are some of the poorest Portuguese regions, have the highest Twitter iPhone usage.

Fig. 12
figure 12

Is iPhone usage a true wealth indicator? a Wealth per capita (100 is the EU average), b iPhone Twitter usage

Full size image

5.4 User age

Social networks such as Facebook are currently used by all age groups. Nowadays, more than 60 % of the US Facebook users are over 35 years old.Footnote 6 This also applies to Facebook users in PortugalFootnote 7 (and probably is generalized). The same applies to Twitter users in the USA. In fact, the percentage of users in each age group is very similar between Facebook and Twitter users in the USA (Fig. 13). However, it is not known if a similar age distribution also applies to Twitter users in Portugal. This is due to the fact that, until October 2015, Twitter registration did not include age information (not even as an optional field). Due to this fact, Twitter age demographics have been obtained mostly using traditional surveys. In Portugal (and in many other countries), no such survey was conducted (to our knowledge) due to the relatively low Twitter usage compared to Facebook. Methods to automatically extract Twitter user age indication based on user activity and user profile are possible, but not necessarily reliable, and have not been applied to Portuguese user datasets.

Fig. 13
figure 13

Percentage of users in each age group for Facebook and Twitter in the US

Full size image

Given the above facts, and in the absence of better information, it could be assumed that Twitter Portuguese user age distribution is similar to Facebook. However, a small experiment performed on our localized database shows that such assumption must be very wrong. Motivated by an earlier work (Brogueira et al. 2014b), we looked for word trigrams within the tweets contents and found out that the top 2 trigrams are:

  1. 1.

    “a minha mãe” (my mother);

  2. 2.

    “o meu pai” (my father).

The combined frequence of those 2 trigrams is three times larger than the next trigrams: “o que é” (what is); “com a minha” (with my). Among the top list there are also many trigrams involving the words “school” and “class”, which in Portuguese are only used up to high school level of education (in higher education levels different names are used, e.g., the equivalent to “university” instead of “school”). We believe, even without further evidence, that this is a strong indication of a very young user base age, and that, contrary to USA, a very large percentage of the geolocated Twitter users in Portugal are teenagers.

This fact is supported by most of the results shown in the previous sections and shows that some demographic indicators concerning social networks cannot be inferred form other countries or studies involving other social networks.

6 Case study example: Portuguese tourists in the Algarve

The Algarve is the southernmost region of continental Portugal. It has an area of 4997 km2 with over 450,000 permanent inhabitants, the Algarve is the most popular tourist destination in Portugal, and one of the most popular in Europe. Its population triples to nearly 1.5 million people in the peak holiday season thanks to seasonal residents, and received more than 16 million bookings (night stays) in 2015,Footnote 8 of which 12.7 million corresponding to 2.6 million foreign tourists. From the official numbers, one can deduce that the Algarve receives around 3.3 million bookings from around 1 million tourists. However, such numbers are a much less precise estimation than the numbers accounted for foreign tourists. As we will see, this is due to the way on how such numbers are accounted for, and the specificities of the local tourism.

The number of tourists is usually accounted based on the number of night stays by nonlocal residents. In the case of foreign tourists in the Algarve, such numbers are fairly accurate since most tourists spend their nights at officially accounted locations, ranging from camping sites to luxury villas, passing by the most diverse range of accommodations (hostels, guesthouses, aparthotels, high rise resorts, hotels, etc.). Even locally owned apartments rented to foreigners on websites such as Booking.com are officially accounted for.

However, such method does not hold for national visitors since many Portuguese citizens residing in the larger urban areas (Lisbon, Oporto, etc.) own houses or apartments in the Algarve, and an even more significant number of Portuguese citizens traditionally spend part of their summer vacations in the Algarve in apartments or rooms owned by friends or locals. These tourists are usually not accounted for by the official methods since their stays are not registered: Owners do not state when they are occupying their houses; rentals are usually not officially declared in order to avoid paying taxes (clients save more than 20 % VAT; owners avoid paying income tax ranging from 20 % to almost 50 %). Usually, these rentals work on a mouth-to-mouth basis, and most people return the following year and have been doing it since their parents’ time. This tradition holds since the 70s, when with the end of the dictatorship regimen, families from all social strata were able to start vacationing in the Algarve (the Algarve had become a very popular destination for northern European tourists earlier in the 60s). Most families from the big cities would try to spend at least 1 week in Algarve despite their income. Lower classes were not able to afford traditional lodging, and locals started renting bedrooms (some even their whole houses) during the traditional summer vacationing period (mid-July to mid-September) in order to improve their income. Inhabitants of smaller villages started building small annexes or adding extra rooms that would satisfy the demand and cater for all sort of clients’ income. Later, during the 80s construction boom, locals even started investing in buying apartments for rental. As mentioned above, such short-term rentals (usually on a weekly basis, from Saturday to Saturday) have never been made official despite the government effort, and even nowadays, only web-based reservations (usually weekends during the off season since the summer period is guaranteed since the previous year) are declared. As such, it is very difficult to obtain national tourist statistics in Algarve during the summer period except by recurring to traditional surveys.

In this section we exemplify how the developed work and the use of geolocated tweets can be useful in future works concerning the characterization of the Portuguese administrative regions, by trying to automatically identify the origin of Portuguese tourists in the Algarve during the summer period, and most important, trying to automatically detect how long they stay in the region.

We considered the period from June 15 to September 15, and started by detecting how many users that spend the rest of the year tweeting from a different region have tweeted in the Algarve during that period, obtaining in the process a database of possible Portuguese tourists in Algarve. Figure 14 shows the number of tweets per day by Portuguese tourists in the Algarve during the considered time span. It is possible to see that August is still the preferred vacations period, which is justified by the fact that this is the period when basically all education-related facilities are closed in Portugal (from Universities to nurseries), forcing families to choose this period for their vacations. Note that facilities associated with the mandatory education range (6–17 years old) are usually closed from mid-July to mid-September, explaining the also relevant number of tweets by tourists in Algarve during that range.

Fig. 14
figure 14

Tweets per day by Portuguese tourists in the Algarve during the summer period

Full size image

Then, we focused on the region of origin of those users. Figure 15 shows the users region of origin by 2-week periods. Darker tones indicate a larger number of users. It is possible to see that the distribution per region of origin is roughly constant during the considered period and that, as expected, more densely populated regions contribute with more tourists. An interesting exception is the district of Braga, which despite being the fourth most populous district, contributes with an unusually low number of tourists (throughout all summer).

Fig. 15
figure 15

Origin of Portuguese tourists in Algarve during summer 2014 (from left-top to bottom-right): a Jun 15th–30th, b Jul 1st–14th, c Jul 15th–31st, d Aug 1st–14th, e Aug 15th–31st, f Sep 1st–14th

Full size image

Finally, we tried to detect how long each user spends in the Algarve region. For each user, we considered only uninterrupted tweeting sessions while in the Algarve (i.e., the user must not have tweeted from a different region between the first and the last Algarve tweet.) Unfortunately, the number of users that tweets everyday is very low among the considered Portuguese tourists in the Algarve, and as such it is not possible to simply consider the difference between the dates of the first and the last tweet while there, since it is not guaranteed that the user tweeted on arrival and on the last day. This difference can only be used to indicate the minimum length of stay. It is also not possible to consider last and first tweets in the region of origin, since the user might have stayed in the region origin without tweeting—this value can give at most an indication of the maximum stay in Algarve (which is the date difference −2). For each user, we considered the minimum and maximum possible lengths of stay, and selected those users whose difference between maximum and minimum length of stay is 2 or less. This way we can guarantee that the user length of stay in the Algarve has an error of at most 2 days. We obtained a database consisting of almost 2 K users satisfying this condition.

Based on the previous values we built an histogram of the length of stay of Portuguese tourists in the Algarve during the summer (Fig. 16). Due to the uncertainty in the length of stay of each user, each column in the histogram has an associated uncertainty related with its immediate neighbor columns. Nevertheless, the uncertainty, the histogram reveals some very interesting data that is consistent with the perception of Portuguese tourists stays in Algarve.

Fig. 16
figure 16

Histogram for the length of stay of Portuguese tourists in Algarve during summer 2014 (with a 2-day error margin)

Full size image

The most usual length of stay is 7 days, which is consistent with the perception that Portuguese tourists maintain the tradition of spending their Algarve vacations on a weekly basis. If one considers the neighbor values (6- and 8-day stays) and the considered uncertainty, this perception is even more reinforced. When considering stays longer than 1 week, it is possible to see a local maximum at the 2-week mark, which is also consistent with the fact that wealthier families rent for 2 weeks instead of just one. The shape of the histogram and its uncertainty certainly allows us to assume that most values indicating stays between 1 and 2 weeks are residual and correspond to either 1- or 2-week stays.

There is a relevant number of shorter stays (2–4 days) that probably correspond to weekends and long/extended weekend hotel stays (Friday to Sunday, Friday to Monday, etc.), and probably mostly in June–July. Further studies focusing on smaller time periods could confirm/clarify this hypothesis.

It is also possible to observe stays longer than 2 weeks that probably mostly concern stays in own properties, e.g., house owners’ children spending their summer school vacations in Algarve and retired users.

7 Conclusions and future work

This paper presents an analysis over a database of about 27.8 Million Portuguese geolocated tweets, produced in Portugal by 97.8 K users during a 1-year period. By observing the geolocated Twitter usage by the Portuguese community, this paper reveals that it is possible to extract relevant indicators such as: the daily periods of increased activity; the prediction of regions where the concentration of the population will be higher or lower in certain periods of the year; what are the most satisfied and dissatisfied regions; what Portuguese use Twitter for; what do Portuguese tweet about; and who are the Twitter users. Such information could prove useful for areas as different as marketing, tourism, sociological studies or even public health.

The decision to base the study solely on geolocated tweets has the advantage of allowing us to remove the influence of most tweet mass producing users, which tend to distort statistics due to their weight within the twitter community.

Among the most interesting conclusions is the fact that the Portuguese community on Twitter must be in large part constituted by teenagers that uses this social network essentially to chat, and exchange thoughts to friends, instead of news dissemination. Moreover, the most discussed topics involve sports and tv shows, instead of “more serious subjects”. It was also possible to denote the negativism and frustration usually associated with the Portuguese people, and the notorious absence of anger and revolt.

This paper is a first step in understanding the idiosyncrasies of Portugal and the Portuguese regions in terms of contents in daily-based or yearly-based periods. The presented analysis, included a small case study of Portuguese tourists in Algarve, shows just a few examples of what can be done with the available data. This study will be further extended in order to better characterize each of the regions in terms of daily habits, user profiles, and also in order to better understand the way people travel across regions.