1 Introduction

The topic of vaccination has attracted considerable controversy over the years. Since the very beginning of widespread vaccination in the early 1800s, there have been groups of people actively opposing their inoculation (Wolfe and Sharp 2002). Their resistance (which sometimes escalated into riots) leads to the foundation of opposition movements: the “Anti-Vaccination League” was first founded in London in 1853, followed in 1867 by the “Anti-Compulsory Vaccination League,” which only focused on the infringement of personal liberty and choice due to vaccines being mandatory (Wolfe and Sharp 2002).

As of today, the belief that vaccinations possibly lead to severe adverse effects is still popular. A recent worldwide survey (Larson et al. 2016) shows that despite the majority of the people being ultimately aware of the benefits of vaccination programs, the percentage of the population raising doubts on their safety or actively opposing to their inoculation is significant and is growing. Researches are alarmed by this situation and worried about public health being at risk, as they underline that the loss of public confidence may lead to the decrease in immunization rates and to the resurgence of diseases (Bonhoeffer and Heininger 2007; Chen 1999). Official health organizations further back this alarm, as they have registered outbreaks of infectious diseases (e.g., measles) in recent years in both Europe and the USA (Zipprich et al. 2015; World Health Organization et al. 2018).

In Italy, the debate on vaccinations experienced has intensified significantly in the last few years. The decreasing vaccine coverage rates and the re-emerging measles outbreaks threatened to compromise the so-called herd immunity (Signorelli et al. 2017b, a). The Italian parliament reacted by promulgating a law that introduces the obligation of ten mandatory vaccinations for infants and their exclusion from kindergartens if they are not in compliance with all the required vaccinations (Donzelli and Demicheli 2018).Footnote 1 This resulted in a widespread controversy, as a free-vax movement has emerged to oppose against the new law and contest its compliance with the Italian Constitution.Footnote 2 As a result, the media coverage and the public participation on social media on this topic have been particularly wide (Furini and Menegoni 2018).

With the advent of social networks, the debate on vaccinations found fertile ground to grow on—not only in Italy but on a worldwide basis (Bello-Orgaz et al. 2017). Researches recognize that the anti-vaccination groups have successfully exploited the power of the Web 2.0 and proactively push medical institution into providing more comprehensive and convincing information concerning the safety of vaccines (Betsch et al. 2012). It has been proven that the exposure to proper Web-based resources can positively influence parental vaccine behaviors (Glanz et al. 2017; Biasio et al. 2016), but the spreading of fake news can create a great deal of confusion on the subject (Ciampaglia 2018), as confirmed by a recent work that studied the issue of vaccine hesitancy in Italy (Giambi et al. 2018). In any case, social networks are certainly playing a central role in the debate and in the campaigns of activists that either advocate for or oppose to vaccines.

This scenario motivated us to gain a deeper knowledge of this phenomenon. The study that we present in this paper was first funded in 2015 and 2016 by the health department of an Italian region (Veneto) and continued as a long-term independent research work. Here, we adopt a Social Business Intelligence approach (Francia et al. 2016) to systematically monitor the active discussion on the theme of vaccinations in Italy: for a period of over 3 years, we have collected all social media posts, blog entries, and Web pages that mentioned this subject.

In this paper, we start with a general overview of the discussion, analyzing trends, occurrences, and correlations of specific topics. In particular, we rely on domain experts from the health department for the definition of the topics of interest that need to be monitored. Then, we go more in depth to analyze the communities that have been formed on Twitter around the discussion of vaccinations from different perspectives. We adopt two distinct approaches to retrieve both active communities (i.e., users that actively participate in the discussion by commenting on vaccinations) and passive communities (i.e., users that read comments from the most influential ones), and we characterize them based on the emerging opinion class (i.e., either in favor of vaccine, against vaccines, or against the mandatoriness of vaccines). Finally, we infer knowledge of these communities based from both structural and semantic points of view.

Differently from previous works on this subject:

  • we adopt an approach that enables a long-term and continuous monitoring of the discussion, and we cover a wide time span on a variety of media types;

  • the distinction of active and passive communities enables a deeper understanding of the discussion and of the involved users;

  • to best of our knowledge, we are the first to consider the free-vax classification (i.e., those that oppose only to the mandatoriness of vaccines) in the definition of communities. This allows us to gather interesting information on a movement that has grown with the introduction of the mandatoriness law.

The paper is structured as follows. Section 2 describes the related works on the theme of vaccinations, while Sect. 3 discusses the methodology we adopted in this study; the results of the analyses are then presented in Sect. 4. We conclude with final remarks and a discussion of future directions in Sect. 5.

2 Related works

The analytical potential hidden in social data has been proven so far by several studies on different domains [e.g., visual impairment campaign (Al Zayer and Gunes 2018), political polarization (Conover et al. 2011; Francia et al. 2016; Trottier and Fuchs 2014), social behaviors (Holmberg and Thelwall 2014), brand analysis (Ghiassi et al. 2013), just to name a few], reaching the highest attention by the mainstream media and population with the recent Cambridge Analytica scandal (Grassegger and Krogerus 2017). A whole research area (called Social Business Intelligence (Francia et al. 2014), or SBI) is dedicated to the integration of the social data flow into the enterprise Business Intelligence pipeline to improve the analytical potential of managers and decision makers (Gallinucci et al. 2015). Thus, it is no surprise that many studies have relied on social data to better understand the discussion on vaccinations and to acquire the necessary knowledge to improve the efficacy of official communication channels (Radzikowski et al. 2016).

A summary of the related work is given in Table 1, where the last line represents our work. Since most research is focused on Twitter, we discuss Twitter-based works in Sect. 2.1 and the others in Sect. 2.2.

Table 1 Summary of related works
Full size table

2.1 Works on Twitter

Among Twitter-based studies, the most popular activity is sentiment analysis, which consists in detecting the polarization (either positive, neutral or negative) expressed in a post by its author; this is usually done by relying on natural language processing (NLP) techniques that infer the sentiment by analyzing the morphology, syntax, and semantics of the text. Although several tools exist for automatic detection of sentiment, related works build their own classifier, train it on a manually labeled sample, and use it to label the whole corpus. D’Andrea et al. (2017) focus on Italian discussion and project the obtained sentiment values on a timeline to understand whether specific events (e.g., the cancelation of the “Vaxxed” movie projection at the Senate) caused a spike in either negative or positive tweets. Dunn et al. (2015) also focus on sentiment trends to demonstrate that the likelihood of a user posting a negative tweet after exposure to a majority of negative opinions is higher than those exposed to a majority of positive and neutral tweets. Other works (Kadam 2017; Mitra et al. 2016; Salathé and Khandelwal 2011) rely on labeled tweets as a starting point to infer the polarization of users and to carry out further analyses. Kadam (2017) runs text analytics on the groups of polarized users and builds a classifier to predict a user’s orientation based on the content of her tweets. Mitra et al. (2016) further distinguishes between active users that steadily tweet either in favor or against vaccinations and “sleeping” users that tweet against vaccinations only after a significant event. Linguistic inquiry and word count (LIWC) (Tausczik and Pennebaker 2010) and meaning extraction method (Chung and Pennebaker 2008) text analysis techniques are used to understand social and behavioral characteristics of each group, concluding for instance that anti-vaccination groups tend toward categorical thinking and conspiratorial worldviews, or that they show a higher group cohesion. Salathé and Khandelwal (2011) used the polarization of users to infer the polarization of communities. Starting from the dataset of tweets (focused on the influenza A(H1N1) vaccine delivered in 2009), they build the network of the tweeting users and of their following relationships and run the Spin Glass community detection algorithm (Reichardt and Bornholdt 2006) to discover the existing communities. Ultimately, they conclude that almost all of the significant communities are polarized and that users tend to seek information from those sharing the same opinion (a behavior that increases the risk of an outbreak among nonprotected individuals).

Community detection based on the following relationships is also done by Surian et al. (2016), which uses Louvain (Blondel et al. 2008) and InfoMap (Rosvall and Bergstrom 2008) community detection algorithms and focuses on tweets about the human papillomavirus (HPV) from 2013 to 2015. Here, instead of sentiment analysis, they use latent Dirichlet allocation (LDA) (Blei et al. 2003) and Dirichlet mixture model (DMM) (Nigam et al. 2000) text analysis techniques to obtain (and manually label as positive or negative) the topics mentioned in the tweets; then, they infer the polarization of communities by verifying the prevalence of either kind of topics. Other than the typical pro- and anti-communities, they discover an “experiential” community, consisting of young people tweeting about their experience with vaccines which are at greater risk of exposure to safety concerns.

Other works (Yuan and Crooks 2018; Bello-Orgaz et al. 2017) run community detection algorithms on the network based on retweeting relationships (i.e., a user is linked to another if the first retweeted a tweet from the second one) rather than following relationships [although Surian et al. (2016) claim that the latter yields the highest levels of performance]. Similarly to Salathé and Khandelwal (2011), Yuan and Crooks (2018) use the sentiment detected on tweets to infer the polarization of the communities, identified with the Louvain algorithm (Blondel et al. 2008). By analyzing the communities, they discover that analysis shows that people in the anti-vaccination community tend to communicate between themselves, while those in the pro-vaccination one tend to retweet and respond to anti-vaccination tweets—thus suggesting that studies evaluating the influence of pro-vaccination users should consider whether their tweets actually penetrate the anti-vaccination community. Bello-Orgaz et al. (2017) run a comparative evaluation of community detection algorithms (concluding that Fast-Greedy (Clauset et al. 2004) is the one yielding the most cohesive and dense communities) and infer the polarization of communities by manually evaluating the most frequent retweets in the community. Their conclusions are that anti-vaccination communities tend to be less in number and less connected than pro-vaccination ones and that famous people tend to fall in the latter groups.

A different approach is taken by Kang et al. (2017), which focused on the links to external Web sites posted in the tweets: they identify the most popular links, manually label them as positive or negative based on their actual content and transcribe the latter into subject–predicate–object triples, so as to create a semantic network of concepts. Their results show that the pro-vaccination group is more cohesive in terms of the topics and discourses appearing in the linked Web sites.

Radzikowski et al. (2016) focuses instead on hashtags, discovering the most tweeted ones and building a network of their co-occurrences. Their findings show that news stories about health issues are the ones driving public participation, but official public health agencies are not strongly featured in the retweet narrative. In particular, bottom-up campaigns and grass-roots activism far outweigh the impact of top-down efforts from authoritative sources such as CDC and WHO, thus suggesting that an indirect approach from mainstream media may be more effective than a direct approach from governmental agencies.

2.2 Works on other media sources

Although Twitter captures most of the researchers’ attention, there exist works based on different sources. Furini and Menegoni (2018) rely on Facebook to analyze the discussion in Italy: the texts of posts from over 200,000 groups (manually labeled as anti- or pro-) are analyzed to identify the presence of words belonging to some predefined categories (affective, social, medial and biological) and subcategories. Their results indicate that pro-vaccination groups show more anxiety and talk more about health in general, whereas anti-vaccination groups show more anger and talk more about vaccine damages to the human body. Another work on Facebook is Faasse et al. (2016), which uses LIWC to analyze the responses to a popular post on vaccines. Its results indicate that despite the absence of scientific evidence, anti-vaccination people show greater analytical thinking; this suggests that their stance is not “universally critical” of vaccines and that it may originate from differing understandings of the benefits and risks of vaccination.

Ultimately, few works fall outside the scope of social networks. Covolo et al. (2017) focus on Italian YouTube videos to discover that anti-vaccination videos are the most shared and liked ones despite being fewer in number. Larson et al. (2013) suggest an architecture for real-time statistical analysis of content from a variety of sources, including articles, blog posts, and news stories, in order to tailor more effective and timely strategies for immunization programs. Although interesting, their work still relies greatly on manual tasks. Yom-Tov and Fernandez-Luque (2014) take a distinguishing approach and focus on search queries on Bing to understand the behavior of pro- and anti-vaccination users. Findings show that, when looking for information on vaccinations, users on both sides are biased in issuing their queries and that although they ultimately look for the same kind of information, their browsing behavior reacts differently after reading the same content. Interestingly, the reading of extreme opinions always proved to have repulsive effects, steering the user’s browsing behavior to more moderate and less opinionated content.

3 Methodology

An overview of the approach we adopted is provided in Fig. 1. The core part is compliant with the SBI methodology presented in Francia et al. (2014) and is central to the analyses described in this paper. In particular, we distinguish between the topic-driven analysis, carried out on the contents crawled in the observed period, and the network-driven analysis, which focuses on the communities that have formed around the topic of vaccines.

Fig. 1
figure 1

Approach overview

Full size image

3.1 The SBI core

The first step in the SBI methodology is the crawling process, required to populate our database with clips collected from the Web. With the term clip, we refer to any kind of textual content that comes in the form of a Facebook post, a tweet, a comment on a blog or a news Web site, an article, etc. The retrieval of clips requires the adoption of a certain technique to filter only those related to the specific topic of interest. To this end, argumentation mining has emerged as an interesting approach to automatically extract structured arguments from textual documents (e.g., Lippi and Torroni 2016; Habernal and Gurevych 2017). In this research project, we relied on the commercial software Brandwatch (www.brandwatch.com), i.e., a social media monitoring tool for the analysis of user-generated content. It provides a keyword-based crawling service to capture the clips whose content is matched by a certain set of queries. In particular, we issued a single query looking for Italian clips containing any conjugations of the term “vaccine.” The query included filters to exclude false positive: most importantly, it discards any mention of animal-related vaccinations and it excludes Web sites known for simply re-posting content from other Web sites.

Clips are downloaded from Brandwatch on a daily basis. As they are stored in the database, a preprocessing step called semantic enrichment is taken to run a preliminary textual analysis on the clips’ content. The goal is to dissect the textual content into single words so that the ones that have been used the most can be detected. Consistently with the SBI methodology (Francia et al. 2014), we consulted with experts in the local health department of the Italian region of Veneto to define an ontology of topics that they felt relevant to be monitored; we refer to it as the Topic Ontology. We started by identifying a set of themes of interest to the experts (e.g., vaccine types, fears, subjects), and we defined the list of topics to be monitored (e.g., “trivalent,” “autism,” “medic”); each topic comprises a list of different aliases that could be used (e.g., “MMR” and “doctor” are aliases of “trivalent” and “medic,” respectively). The Topic Ontology has been maintained throughout the years; Table 2 shows the overall themes identified and the number of defined topics, including some examples. The semantic enrichment process starts by processing the textual content with the open source Apache Lucene library and reducing them to their lemmas (i.e., their canonical form), in order to resolve plurals, masculine and feminine variations; stop words are ignored, while hashtags (i.e., words beginning with the “#” character) are not lemmatized. Then, words are linked to the topics in the Topic Ontology.

An overview of the collected dataset is given in Sect. 4.1.

Table 2 Themes and topics related to vaccinations, identified with domain experts; the table reports English translations of the Italian terms
Full size table

3.2 Topic-driven analysis

The topic-driven analysis takes place on the semantically enriched content and is aimed at understanding which topics and events are driving the discussions around vaccines. This is mainly done on the topics of the Topic Ontology and on hashtags, which are keywords that often describe synthetically the discussed topic and, possibly, even the opinion of the user. A powerful analysis can be done by examining the co-occurrence of different topics among the clips; we say the two topics co-occur in a clip if they are both mentioned in its textual content with no more than 20 words in between them. Co-occurrences can highlight the correlations between different topics: for instance, a public institution can improve the efficiency of its informative campaigns by identifying the kinds of vaccine whose safety is frequently questioned; this can be done by analyzing the co-occurrences between the topics of the Vaccines and Fears theme.

The results of the topic-driven analysis are provided in Sect. 4.2.

3.3 Network-driven analysis

While the topic-driven analysis allows for a general overview of what the conversation around vaccines is about, the network-driven one aims at profiling the people directly or indirectly involved.

3.3.1 Identification and classification of influencers

The core aspect of the network-driven analysis is the identification among the reference population of groups of users that either commonly speak against or in favor of vaccinations, or that are passively subject to certain kinds of opinions. Thus, one of the main aspects consists in labeling users and groups based on their respective opinions. In particular, we discriminate between the following opinion classes:

  • P: Pro-vax, i.e., those speaking in favor of vaccines;

  • A: Anti-vax, i.e., those speaking against vaccines;

  • F: Free-vax, i.e., those fighting against the obligation imposed on the children going to kindergartens, but not fighting against vaccines themselves.Footnote 3

In the influencers extraction step, we define a ground-truth comprising the 50 most influential users for each opinion class. To determine the influence of a user, we rely on the influence measure by Bakshy et al. (2011) (we will refer to it as the Bakshy measure in the remainder of the paper): given a certain user, the influence is computed as a function of the number of her followers and the average number of retweets by her direct followers. Since the Bakshy measure returns discrete values, we introduce a further measure to sort users with the same influence. In particular, we calculate the vaccine follower ratio as the fraction of the user’s followers who belong to our reference dataset. In other words, we state that the higher the fraction of followers who tweet about vaccines, the higher is the user’s importance in the vaccine domain. The classification of users is done by three domain experts with a majority voting system. Given the list of users ordered by descending influence, each expert manually ascertains their opinion until 50 influencers have been determined for each opinion class. We will use I to refer to the set of influencers and \(I_{\mathrm{A}}\), \(I_{\mathrm{F}}\), \(I_{\mathrm{P}}\) to the subsets of influencers in the respective opinion classes.

For the purpose of this analysis, we focus on Twitter alone, as its API service allows for easily obtaining the following relationships between users. (Also, we will show in Sect. 4.1 that Twitter is the most relevant source for vaccine-related discussions.) Since Twitter’s APIs return only current data (e.g., the list of a user’s followers cannot be filtered on a certain date), we limit the network-driven analysis to the last month of our dataset (i.e., August 2018).

A discussion on the classified influencers is given in Sect. 4.3.1.

3.3.2 Active and passive networks

The goal of the network-driven analysis is to understand the characteristics of two kinds of networks: the active network, which comprises people actively discussing about vaccines (because they have tweeted about it at least once in the observed period), and the passive network, which comprised people that simply listen to the subject (because they follow at least one influencer). We will refer to users in the active and passive networks as writers and readers, respectively. In both cases, we aim at grouping people into communities (based on their opinion class) and to evaluate the latter from a structural and a semantic perspective. We will use W and R to refer to communities in the active and passive networks, respectively.

In regard to the active network, we start from the set of people who tweeted about vaccines , collect their relationships through the Twitter APIs (as this information is not provided by our crawling service) and run the well-known Louvain algorithm (Blondel et al. 2008) to detect communities. The Louvain algorithm works similarly to hierarchical clustering, i.e., it groups nodes (and subsequently groups) bottom-up; the goal is to locally maximize the modularity, which measures the density of edges inside communities to edges outside communities. Once the communities W are detected, we use the subset of classified users I to verify whether the opinion classes are well partitioned. Ultimately, the claim is that studying the relationships between the users will reveal their opinion.

As to the passive network, we start from the classified users I and study the characteristics of their neighborhood. In particular, we build a community for each opinion class (i.e., \(R = \{ R_{\mathrm{A}},R_{\mathrm{F}},R_{\mathrm{P}}\}\)) by considering the set of followers of the respective influencers. (Again, we rely on the Twitter APIs to collect the list of the influencers’ followers and the relationships among themselves.) The underlying assumption is that the act of following users with certain opinions augments the probability of the follower to agree with the opinion of the followee. Passive communities differ from active ones in two aspects. First, their cardinality is fixed, i.e., \(|R|=3\), whereas |W| depends on the structure of the network itself and on the algorithm used for community detection. Also, passive communities are overlapped, as users are free to follow influencers from different opinion classes. Thus, we will make reference to the exclusive communities \(R_{\mathrm{A*}}\), \(R_{\mathrm{F*}}\), and \(R_{\mathrm{P*}}\) (i.e., including users that follow influencers from only one opinion class) and to the intersection communities \(R_{\mathrm{AF}}\), \(R_{\mathrm{FP}}\), \(R_{\mathrm{AP}}\), and \(R_{\mathrm{AFP}}\) (i.e., including users that follow influencers from more than one opinion class).

A discussion on the characteristics of the active and passive networks is given in Sects. 4.3.2 and 4.3.3, respectively.

3.3.3 Structural and semantic analyses of the networks

The analysis of the two networks begins with the validation of W by framing the manually ascertained influencers I. The goal is to verify whether the opinion of the influencers can generalize the opinion of the writers. Then, we compare the communities of the two networks from both structural and semantic points of views. In the first case, we look for similarities and differences between communities in terms of different structural metrics, including their size, cohesion, and overlap degree. In the second case, we are interested in understanding which are the main topics of interest in each community, without restricting the scope to the theme of vaccinations. To this end, we exploit Twitter APIs one last time to collect the most recent tweets (up to 3200, according to Twitter’s policy) and we examine the usage that communities make of hashtags (i.e., words preceded by the “#” character), which are frequently used on Twitter to provide the context of the tweet. On the one hand, we calculate the Ruzicka index (Deza and Deza 2006) [i.e., a weighted version of the Jaccard index (Jaccard 1901)] to calculate the pairwise similarity of the communities in terms of the used hashtags; the intuition is that the more two communities use the same hashtags, the more they are similar. On the other hand, we isolate the most distinguishing hashtags for each community and we manually evaluate the existence of any pattern characterizing the communities.

The task of finding distinguishing hashtags resembles the one of finding the most relevant terms within a corpus of documents in information retrieval. Among the most popular techniques to address this task are TF–IDF (term frequency–inverse document frequency) (Salton and McGill 1984) and PMI (pointwise mutual information) (Church and Hanks 1990). The first one multiplies the frequency of a term (TF) by its inverse document frequency (IDF), which quantifies the specificity of a term as an inverse function of the number of documents in which it occurs. The second one quantifies the discrepancy between the probability of the coincidence of two variables (a term and a document) given their joint distribution and their individual distributions, assuming independence. The problem with TF-IDF is that it excludes terms that appear in every document. This works well in a scenario with many documents in order to find rare terms that identify some documents; however, in our case, we want to retrieve popular hashtags that are mainly used (but not exclusive to) by a certain community. PMI partly addresses this issue by not excluding the latter hashtags, but still risks to confer a higher score to those that are exclusive to a community, independently of their absolute frequency. For instance, a hashtag that appears ten out of ten times in a certain community has a higher PMI score than a hashtag that appears 999 out of 1000 times in the same community.

To isolate the distinguishing hashtags, we introduce a variation of TF-IDF that best addresses our scenario. Given a set of communities C, for each hashtag h used by a community \(c \in C\) we compute the term frequency tf(h, c) as in TF-IDF as

$$\begin{aligned} \hbox {tf}(h,c) = \frac{f_{h,c}}{f_{c}} \end{aligned}$$

where \(f_{h,c}\) is the number of times h has been used by the users in c, and \(f_{c}\) is total number of times any hashtag has been used by the users in c. Then, we compute the term frequency ratio tfr(h, c) as

$$\begin{aligned} \hbox {tfr}(h,c) = \frac{\mathrm{tf}(h,c)}{\sum _{c' \in C} \mathrm{tf}(h,c')} \end{aligned}$$

In practice, the term frequency ratio measures the relevance of each hashtag in each community based on the calculated term frequencies. Ultimately, we say that a hashtag h is distinguishing of a community c if tfr\((h,c) > 0.5\) (which is possible for only one community, if any). For the sake of this analysis, we focus on the hashtags that show the highest term frequencies and that satisfy the previous condition.

The results of the structural and semantic comparisons are provided in Sects. 4.3.4 and 4.3.5, respectively. In the latter, both PMI and our new approach are used to evaluate the distinguishing hashtags.

4 Results

We first provide a description of our dataset in Sect. 4.1 and then show the results of the topic-driven and network-driven analyses in Sects. 4.2 and  4.3, respectively.

4.1 Dataset description

Fig. 2
figure 2

Monthly trend of clips retrieved on vaccinations; peak periods (letters A to D) are highlighted with different colors and discussed in Sect. 4.1

Full size image

Our dataset covers a period of over 3 years, spanning from October 2015 to August 2018 and containing a total of 2,207,167 clips. Figure 2 shows the monthly trend of clips over the listening period. There are several peak periods that are highlighted in figure, namely:

  1. (A)

    October 2015 is due to first warnings from the World Health Organization (WHO) concerning the low coverage for infant’s vaccinations in Italy.

  2. (B)

    In April 2017 begin the talks of the mandatory regulation, which is approved in July 2017.

  3. (C)

    The spike in January 2018 is caused by political talks hinting to a possible weakening of the mandatory regulation.

  4. (D)

    August 2018 is due to reactions on the promulgation of a law (coming from the newer government) that suspended the mandatoriness for kindergarten enrollments.

Among these events, the second one (B) is the most critical one, as it triggers a significant increase in the talks about vaccinations that becomes structural. Before that time, vaccinations were a niche topic: the discussion was scarce (around 700 posts/day on average) and characterized by occasional and short-lasting spikes (no more than 6000 posts/day) due to some news published by media (e.g., the death of a child, the release of the Vaxxed movie, public statements by important politicians or influencers). After that, the talks on the mandatory regulation pushed the topic to a higher level, as it became a political issue frequently debated by the parties and echoed by the media. Indeed, the discussion has increased by an order of magnitude, with more than 5000 posts/day and spikes of around 50,000 posts/day.

Table 3 Distribution of the collected clips on the different media channels
Full size table

Table 3 shows the distribution of clips across different media channels. Twitter emerges as the most relevant channel despite holding a low market share in Italy (Statista 2018b); this is most probably due to its default policy of making tweets public (unless specified otherwise by the user), whereas the discussion on Facebook [which still retains the highest market share (Statista 2018a)] often takes place within private profiles or protected groups.

4.2 Topic analysis

Fig. 3
figure 3

Boxplot of the monthly occurrences of topics in the Topic Ontology over the whole dataset, grouped by theme

Full size image
Table 4 Ten most mentioned topics, their total number of occurrences in the dataset and the % of clips in which they occur at least once
Full size table

Figure 3 gives a summary of the monthly presence of the Topic Ontology’s themes over the whole dataset; the boxplot also describes the (in)stability of each theme, as the amount of mentions depends on the trending discussions among the public opinion. From this analysis, it emerges that subjects and vaccinable diseases are the most present themes. The discussions often revolve around subjects like children (i.e., the main recipients of vaccines), doctors (i.e., those in charge of prescriptions), and parents (i.e., those responsible for the children’s health), all of which appear among the ten most mentioned topics (Table 4). Also, vaccinable diseases like measles and meningococcus are very frequent. The first is debated from all sides, as pro-vaccination people claim it should have been eradicated, anti-vaccination ones argue against the actual dangerousness of the disease, and official health organizations often report its death count and outbreaks. The second is debated for being a dangerous disease which has been excluded from the list of mandatory ones. Other important topics include the obligatoriness of vaccines (which jumped to the top 10 only in May 2017), schools (due to the proposal of exclude nonvaccinated children from primary school and kindergartens), politicians and political parties, in particular Beatrice Lorenzin being the Ministry of Health promulgating the law on obligatoriness, and Matteo Salvini (i.e., the leader of “La Lega”) and M5S (i.e., Five Stars Movement) as the most popular representatives criticizing the law.

Fig. 4
figure 4

Co-occurrences between the topics of themes Vaccine Type (left) and Fear (right)

Full size image

A more in-depth analysis is provided in Fig. 4, which shows the co-occurrences between the topics of themes Vaccine Type (left) and Fear (right); it is a chord diagram generated with D3 (http://d3js.org/), where the size of each chord depends on the number of co-occurrences. Although few correlations cross the boundary between the two themes, interesting conclusions can be derived from this chart. For starters, it is clear that the fear of autism is associated with the trivalent and hexavalent vaccines. This can be a hint for official health organizations to improve the efficiency of their informative campaigns by focusing their efforts on promoting the safety of those two kinds of vaccines. Also, the topic of epidemics is often associated with the anti-poliomyelitis and anti-influenza vaccines; however, this correlation does not indicate an actual fear of the vaccines: a closer look to the involved clips reveals that the anti-poliomyelitis is cited for being capable of stopping one of the latest great outbreaks, while the anti-influenza is advertised to contain the epidemic that spreads every year. Still on the anti-influenza, its correlation with pregnancy is due to the former one being particularly recommended to weak or exposed subjects, such as pregnant women. Ultimately, it is worth mentioning that the absence of any correlation from mercury and aluminum to any of the vaccine types suggests that the fear of the former is more generic; for this reason, official health organizations may need to design a communicative strategy that tackles these topics in a different way than with autism.

4.3 Network analysis

The results of the network analysis are provided in Sects. 4.3.1–4.3.5. To simplify the reading, we summarize in Table 5 the symbols used to reference networks and communities.

Table 5 Summary of the symbols used to reference networks and communities
Full size table

4.3.1 The classified influencers

A summary of the manually classified users I is shown in Table 6, where users are grouped into four levels based on their influence value of the Bakshy measure: high, medium, low or very low. This reveals a substantial difference between the three classes, which resembles the observations made in Bello-Orgaz et al. (2017). On the one side, the most important anti-vax influencers users are not particularly influencing, as the vast majority of them detain the lowest levels of influence; their set is mainly made of ordinary people, and there is not a single celebrity or famous person among them. On the opposite side, all pro-vax influencers fall within the highest level influence; here, not only ordinary people can be found, but also well-known journalists, politicians, and medics. In particular, the “front man” is Roberto Burioni (Wikipedia 2019), an immunologist that has become famous for his battle in favor of vaccines. Interestingly, pro-vax influencers also include several satirical blogs. The free-vax influencers occupy the middle ground, both in terms of influence level (almost perfectly balanced) and number of famous people, mainly bloggers and politicians. Another interesting evaluation can be done from the political perspective: free-vax influencers mainly fall within the sphere of the Five Star Movement, pro-vax ones mainly belong to the left-wing side, and anti-vax ones hardly show any political affiliation.Footnote 4

Table 6 Influence level of the top 50 influencers manually identified for each opinion class
Full size table

4.3.2 The active network

Over the period observed in this analysis (i.e., August 2018), the dataset presents a total of 66,645 users that had tweeted on the topic of vaccinations. The application of Louvain’s community detection algorithm on these users produces a total of 20 communities with a modularity valueFootnote 5 of 0.4154. Details of the communities are reported in Table 7, while Fig. 5 provides a graphical representation (result obtained using the igraph Python library, https://igraph.org/python/); in the figure, colored squares represent the classified users. Interestingly, the major communities that arise provide a clear partitioning of the influencers based on their classification; thus, we have named the communities based on the opinion class of the contained influencers (U stands for unclassified). All anti-vax influencers and almost all free-vax influencers are concentrated in \(W_\mathrm{AF}\); conversely, pro-vax influencers characterize three different communities, namely \(W_{\mathrm{P}_2}\) (the biggest one, containing the influencers with the highest media exposure, including Roberto Burioni), \(W_{\mathrm{P}_3}\) (a strong community including the majority of the influencers), and \(W_{\mathrm{P}_1}\) (a weaker community, including only few influencers). These statistics alone suggest that:

  1. 1.

    Pro-vax users are higher in number and spread in different communities.

  2. 2.

    There is a net separation between the anti-vax and the pro-vax world.

  3. 3.

    The free-vax world shares much more with the anti-vax world than with the pro-vax one.

In the remainder of the paper, we will use \(W_{\mathrm{P}}\) to refer to \(\{ W_{\mathrm{P}_1} \cup W_{\mathrm{P}_2} \cup W_{\mathrm{P}_3} \}\).

Table 7 Statistics on the 20 communities detected in the active network, including the number of users, the percentage with respect to the total, and the number of enclosed anti-, free-, and pro-vax influencers
Full size table
Fig. 5
figure 5

Graphical representation of the communities detected in the active network; dots represent users, and colored squares represent users \(\in\) I

Full size image

4.3.3 The passive network

Starting from I, we have collected all their followers R (\(|R| = 472{,}047\)) and all the following relationships that occur between them. These users are assigned to \(R_{\mathrm{A}}\), \(R_{\mathrm{F}}\), and \(R_{\mathrm{P}}\) depending on whether they follow some of the classified users in the respective class. Details of each community are reported in Table 8, while Fig. 6 provides a graphical representation; in the figure, each circle corresponds to a community in Table 8 and its size is proportional to the number of users. These data show that:

  1. 1.

    There is little overlapping between the pro-vax and anti-vax communities, i.e., very few people are following both pro- and anti-vax influencers. This suggests that the campaign of pro-vax influencers in favor of vaccinations is probably going to have little impact on the anti-vax community.

  2. 2.

    The anti-vax community is actually small if compared to the others, which is due to the absence of truly influential anti-vax users.

  3. 3.

    An unexpected result is the greater size of the free-vax community with respect to the pro-vax one. This may be explained by the higher presence of Five Star Movement politicians among the top free-vax influencer compared to the presence of left-party politicians among the top pro-vax influencers.

Table 8 Statistics on the communities detected in the passive network, including the number of users and the percentage with respect to the total
Full size table
Fig. 6
figure 6

Graphical representation of the communities detected in the passive network

Full size image

4.3.4 Structural comparison of the two networks

Table 9 Metrics to compare the detected communities, including the number of users (|c|), the number of following relationships such as users (intra(c)), the density (dens(c)), the average centrality degree (centr(c)), the percentage of writers from either \(W_{\mathrm{AF}}\) or \(W_{\mathrm{P}}\) that also belongs to c, and the percentage of users of c that are also writers
Full size table

A comparison of the active network W and the passive network R is shown in Table 9. For each community c, we show the number of users (|c|), the number of following relationships between the users of such community (intra(c)), and two measures of cohesion (Wasserman and Faust 1994), i.e., the density (dens(c)), and the average centrality degree (centr(c)). Density is calculated as the number of relationships against all possible relationships (i.e., dens\((c) = \frac{\mathrm{intra}(c)}{|c|*|c-1|}\)), while the average centrality degree is the average number of relationships for each user \(u \in c\) such that u follows another \(u' \in c\). The table also shows the degree of overlap between the communities: for each passive community c, it shows the percentage of writers from either \(W_{\mathrm{AF}}\) or \(W_{\mathrm{P}}\) that also belongs to c. Conversely, the last column shows the percentage of users of c that are also writers. We summarize our findings as follows.

  1. 1.

    We note that 89% of writers are also present as readers; this means that the vast majority of users tweeting on vaccinations are also followers of at least one of the top 150 influencers.

  2. 2.

    In terms of density, we observe that the cohesion of the active communities is higher than in passive ones; this seems intuitive, as writers are actively talking about the subject of this study, while readers are mostly spectators of the debate (as shown by the overlapping percentages). Interestingly, the most cohesive active communities are \(W_{4}\) and \(W_{10}\), i.e., those that contain the highest number of influencers (as seen in Table 7).

  3. 3.

    When looking at the exclusive and overlapping passive communities, we notice that the cohesion measures progressively increase when moving from the former to the latter. Indeed, \(R_\mathrm{AFP}\) is the one with the highest density and average centrality degree, even with respect to the active communities. This can be explained by the fact that the more users are open to interacting with influencers of different sides, the more interconnected they become.

  4. 4.

    The overlapping percentages seem to confirm the classification of the communities, as 46% of users in \(W_\mathrm{P}\) fall within \(R_\mathrm{P*}\) and 41% of users in the \(W_\mathrm{AF}\) fall within \(R_\mathrm{AF}\). This also strengthens the suggestion that there is a high affiliation between anti-vax and free-vax users.

4.3.5 Semantic comparison of the two networks

Besides comparing communities from a structural perspective, we conclude by comparing them from a semantic point of view. The selection of the most distinguishing hashtags (based on both PMI and our approach, presented in Sect. 3.3.3) depends on which communities are taken into account. Thus, given C the set of considered communities, we evaluate the four scenarios described in Table 10. Scenarios \({\hbox {S}_1}\) and \({\hbox {S}_2}\) differ in that the three pro-vax active communities are considered separately and together, respectively; scenario \({\hbox {S}_3}\) considers the three big overlapping passive communities, while \({\hbox {S}_4}\) considers the exclusive and the intersection passive communities.

Table 10 Different scenarios considered for the selection of the most distinguishing hashtags; the results are publicly available at Mendeley Data (http://dx.doi.org/10.17632/2hfw99xz44.1)
Full size table

The tables presenting the results are available at Mendeley Data (http://dx.doi.org/10.17632/2hfw99xz44.2). As anticipated in Sect. 3.3.3, PMI is more biased toward the exclusivity of hashtags in a community (to the detriment of the hashtags’ frequency), whereas our approach brings out popular hashtags that are more used by a certain community. Nonetheless, both techniques consistently return results that support the same conclusion in most scenarios. In the discussion of our findings, we first comment the results obtained with our technique and then compare them with those obtained with PMI.

Overall, the presented scenarios reveal that one of the main topics of discussion (besides vaccines) is politics, especially in the active community. Before commenting on the results, we clarify that, at the time of the analysis, the ruling government is the one formed on June 1, 2018 by the Five Star Movement (5SM, a big-tent party) and Lega (a right-wing party); these are the parties that have been most critical on the obligatoriness of vaccines, which had been promulgated in 2017 by the Democratic Party (DP, a left-wing party). We summarize our findings as follows.

  1. 1.

    \({\hbox {S}_1}\) reveals a clear characterization of the active communities. Writers in \(W_\mathrm{AF}\) not only show support for the free-vax movement, but they also attack the former Minister of Health and former Prime Minister that promulgated the law on vaccine obligatoriness (i.e., Beatrice Lorenzin and Matteo Renzi, respectively) and their political party (i.e., DP). Conversely, writers in \(W_{{\rm P}_2}\) (i.e., the largest community) are the ones showing the strongest support for the pro-vax movement, while those in \(W_{{\rm P}_3}\) (i.e., the one with the majority of influencers) are mainly critical of the current government; writers in \(W_{{\rm P}_1}\) (i.e., the weaker pro-vax community) cover a middle ground (showing pro-vax support and critics to the government) and often cite satirical Web sites. Hashtags retrieved by PMI confirm these considerations.

  2. 2.

    \({\hbox {S}_2}\) confirms the clear separation between \(W_\mathrm{AF}\) and the pro-vax communities. On the one side, the distinguishing hashtags for \(W_\mathrm{AF}\) are mostly the same as in \({{\rm S}_1}\). On the other side, the distinguishing hashtags that emerge for \(W_{\rm {P}}\) can be traced back to the same topics, i.e., pro-vax support and critics to the government and to free-vax politicians. As in \({\rm{S}_1}\), the same is evident in the PMI results.

  3. 3.

    \({\hbox {S}_3}\) shows other themes besides politics that characterize the passive communities. Readers in \(R_\mathrm{A}\) show support toward the Lega party and interest in political talk shows and finance topics; news events appear as the main topic among readers \(R_\mathrm{F}\), while those in \(R_\mathrm{P}\) mostly discuss reality shows. These considerations are less apparent in the PMI results, where only the interest in news events appears in \(R_\mathrm{F}\).

  4. 4.

    \(\mathrm{S}_4\) shows that the support toward the Lega party and the interest in talk shows can be narrowed down to \(R_\mathrm{AF}\), while the interest in finance topics can be narrowed down to \(R_\mathrm{A*}\). Beyond that, no other main topic seems to characterize the remaining communities. Similarly to \({\hbox {S}}_3\), no main topic of interest is highlighted by PMI.

  5. 5.

    From a linguistic perspective, a surprising result is the higher tendency of the distinct groups to argue against those with opposing views rather than directly supporting the respective view. For instance, in both \({\hbox {P}S_1}\) and \({\hbox {S}_2}\) pro-vax politicians and parties are attacked by the anti- and free-vax communities, while free-vax politicians and parties are attacked by the pro-vax communities. This is probably a symptomatic effect of the social network dialectic, in which it is easier to criticize “enemies” rather than supporting “friends.”

  6. 6.

    The terms “pro-vax” and “no-vax” (the latter referring to the anti-vax movement) appear as distinguishing hashtags in \(W_\mathrm{AF}\) and \(W_\mathrm{P}\), respectively. This shows that such terms are mainly used as pejorative labels by the respective sides rather than for self-identification.

Table 11 Pairwise Ruzicka similarities between communities
Full size table

We conclude the analysis by discussing the pairwise Ruzicka similarity between communities in terms of the most used hashtags; the intuition is that the more two communities use the same hashtags, the more they are similar. The results are presented in Table 11. The comparison is also made with the top influencers in each opinion class, i.e., \(I_\mathrm{A}\), \(I_\mathrm{F}\), and \(I_\mathrm{P}\). We summarize our findings as follows.

  1. 1.

    Among the influencers (which are strongly focused on the theme of vaccination), we observe a high similarity between the anti-vax and the free-vax (which further confirms the strong connection between the two groups), whereas both of them appear quite distant from the pro-vax.

  2. 2.

    This trend is further confirmed in the comparison of active communities: the anti-vax community (i.e., \(W_{4}\)) shares similarities only with \(I_\mathrm{A}\) and \(I_\mathrm{F}\), while all pro-vax communities share similarities with each other and with \(I_\mathrm{P}\).

  3. 3.

    As to the passive communities, we primarily focus on the exclusive ones (i.e., \(R_\mathrm{A}\), \(R_\mathrm{F}\), and \(R_\mathrm{P}\)). Users in these communities are less focused on the theme of vaccination and more active on the mainstream topics; as a result, the widespread usage of mainstream hashtags makes interest areas appearing quite similar to each other and, at the same time, quite far from every influencer group and every active community. Nonetheless, it is still observable how users in \(R_\mathrm{F}\) actually cover a middle ground of topics between those in \(R_\mathrm{A}\) and \(R_\mathrm{P}\).

5 Conclusion

In this paper, we analyzed the social debate on vaccines in Italy. The topic is relevant since in the last few years it passed from being a niche audience to a mainstream argument due to the promulgation of a law that makes a set of ten vaccines obligatory that turned vaccine in a political issue. Differently from previous papers, we integrated several techniques and we considered all the available Web and social sources. This allowed us to carry out a comprehensive analysis that tackled the topic from several different points of view. The main outcomes of the analysis are: (a) the fear of autism is mostly associated with the trivalent and hexavalent vaccines; (b) the anti-vax community is quite smaller than the pro-vax one but is also more cohesive; (c) there is no public figure among the anti-vax influencers, contrarily to the pro- and free-vax ones; (d) the free-vax community shares many similarities with the anti-vax one; and (e) there is a clear characterization of the active communities, especially from a political perspective. Finally, as computer science researchers, we deliberately avoid an evaluation from the sociological and medical perspectives, and we advocate for professionals in these fields to exploit the facts and evidences that have emerged in this study.

We will continue monitoring the vaccine topic with particular interest to the evolution of the network structure and to the topic discussed. In particular, it will be interesting to analyze how emerging positions and opinions (like the one experienced in 2017) impact on the Twitter follower network. This requires the adoption of a dynamic approach that allows the tracking of the evolution of communities and the association of such changes to significant events (Wang et al. 2018). Remarkably, since Twitter APIs do not allow the retrieval of historical relationships (nor do they associate a time reference with such relationships), this is possible only by continuously monitoring the relationships in the community.