Inside Trending Topic Algorithm: How Do Human Interactions Drive Public Opinion in an Artificial Environment

Literature Review

The omnipresence of social structures can be found within the daily behavior of social actors, up to co-occurring in the determination of their own identity (Maretti, 2018; Walther & D’Addario, 2001). In this frame is conceived the figure of the prosumer, the one who simultaneously becomes the producer and user (consumer) of the media product. The prosumer figure finds ample space and is reworked within the digital space (Choi, 2015; Kotler, 2010; Ritzer & Jurgenson, 2008). In fact, it is more difficult to distinguish the difference between producers and consumers in the immaterial and collaborative worlds of Web 2.0, where the hegemony of prosumers is clearer.

The digital prosumer, however, is not a particular figure but defines any user on the network. The action of producing digital content within the web space, in fact, is not limited to the creation of specific online multimedia content (e.g., post, video, or image), but it also concerns simpler and apparently low communicative activities (e.g., a share or retweet, a like or mention or geolocate yourself in a specific place). Even choosing to open one website over another means having also contributed unwittingly to managing it.

In the digital universe of prosumer 2.0 and mass communication (Castells, 2014), a fundamental role is that of the so-called hubs, connectors (Gladwell, 2006), or influencers (Bakshy et al., 2011). Hubs are influential social actors who possess a number of preferential links within affinity networks and are able to influence the rest of the network. In digital social networks, hubs can generate fashion, spread customs, and determine the virality of information.

According to the theories of Barabási (Albert & Barabási, 2002), the action of the hubs within the exchange networks responds to the law of the “rich who becomes ever richer” (i.e., the theory of invariant scale networks; Albert & Barabási, 2002, p. 27). According to this theory, the influencers present within the network will naturally tend to acquire links and, therefore, consent in a greater way than nodes with a lower degree. For this reason, the most popular sites are more visited than the less popular ones and tend to acquire more links than other sites, and blogs with more followers are subject to acquire more contacts than the smaller ones.

In the 1960s, Lazarsfeld et al. (1968) were the first to define the role of specific individuals, called opinion leaders, in influencing public opinion. The peculiarity of these social actors is that their influence is attributable solely to their relational capacity and their exposure to the mass media. The presence of these particular actors redefined and mediated the flow of communication from the medium to the public, which became a two-step flow (Katz, 1957; Katz & Lazarsfeld, 1955). The theory developed by Lazarsfeld (Katz, 1957; Katz & Lazarsfeld, 1955) refers to the world of mass media; otherwise, in the digital space, the role of opinion leaders acquires a different value. In the network, exposure to the means of communication is available to all users, social relations can be mediated by social media, and the communication tools are manifold. Therefore, digital opinion leaders are users who manage to acquire a strategic position in the network to such an extent that their contents become resources capable of attracting the attention of the public and acquiring the power of virality (Choi & Park, 2015; Giurgiu & Barsan, 2008).

The behavior of relevant social actors capable of generating social influence also emerges in the studies on Granovetter’s models of collective behavior (Granovetter & Soong, 1983). According to Granovetter’s threshold model, the emergence of collective movements is based on two conditions: (1) a population of agents with binary choices and (2) the influence of the behavior of others. Therefore, if an individual “instigator” acts autonomously, it will provoke a domino effect such as to generate a macro collective behavior. Opinion leaders have some power over their followers and exercise this power by influencing their behavior (i.e., their choice of action). Therefore, after all the actors have chosen their actions, a decision-making mechanism determines the collective choice, resulting from individual choices but influenced by a relevant social actor (Krassa, 1988; Valente, 1996; van den Brink et al., 2011).

Lazarsfeld’s theory (Katz, 1957; Katz & Lazarsfeld, 1955) becomes the basis for Gladwell’s (2006) studies on the phenomenon of virality. According to Gladwell (2006), there are three factors that lead a phenomenon toward the explosion: (1) the law of the few, (2) the grip factor, and (3) the power of context. According to this theory, the explosion of viral phenomena depends, first of all, on the type of infected people involved, called contaminators. Contaminants must have a large network of connections, be able to create what Merton (1948) and Watzlawick (1984) called self-fulfilling prophecies or be sellers or have a particular charisma that allows them to convince others of their arguments. Opinion leaders must be able to modulate their message in relation to the receiving public and to assess the “fertility” of the context to which the message is being addressed.

Trending Topics (TTs) and Opinion Leaders in the Twittersphere

The polarized structure of the Twittersphere is attributable to both the presence of influencer users and popular hashtags. The popular hashtags are signaled by the TT algorithm, which automatically suggests to the user the topics that are of great interest at a particular time and that will naturally acquire more content by polarizing the topic of discussion (Aiello et al., 2013; Lee et al., 2011). The polarization process also nestles specific users who present the typical characteristics of opinion leaders. In particular, on Twitter, the role of the leader is updated and expanded through continuous interactions of digital social actors (Murthy, 2012, 2018) that build and maintain around the leader a relational network constantly fed via three main tools: follow, retweet, and mention (Bentivegna, 2014, 2015; Cha et al., 2012).

The Research Project

The exposed theoretical framework outlines a series of ontological and phenomenological issues, concerning the development of public opinion in the digital space. First, it is necessary to ask whether the Lazarsfeld paradigm (Katz & Lazarsfeld, 1955) of the two-stage flow of communication still represents a valid point of reference for the social dimension. Second, it is necessary to evaluate the characteristics of opinion leaders (or influencers) within the same structures. Finally, it is necessary to ask what the typical language is of the public within the specific affinity networks, without forgetting that each digital structure present on the web has its own public and informational context and carries a specific form of communication (Rainie & Wellman, 2012). The objective of this contribution is to make operative the issues previously exposed in the Italian Twittersphere. The research design translates into the following research questions and hypotheses:

Research Question 1: If the TTs are the preferential nodes of the hashtag semantic network, which are the other co-occurring elements? Is there a transversality between them?

The Research Stages

A specific analysis technique was used for each research segment.

Research Question 1: If the TTs are the preferential nodes of the hashtag semantic network, which are the other co-occurring elements? Is there a transversality between them?

An analysis of the six semantic networks of each sample aimed at identifying the hashtags that co-occurred most frequently with the main TT with the aid of the Gephi software (Bastian et al., 2009). Finally, the transversality of certain specific elements was detected and a semantic graph was constructed to illustrate the hashtag network and to identify, through centrality measurements, the influential elements present together in the TTs.

Research Question 2: How is the relational structure set up within a hashtag?

The reconstruction of the relational structure of the various topics was carried out by considering the total number of elements (consisting of tweets and retweets) present within each topic in relation to the number of individual tweets. Finally, for each tweet, the retweet frequency was evaluated on a scale: 0−10, 11−100, 101−200, 201−300, and from 300 upward.

Research Question 3: Can the influential social actors be pointed out in the six TTs?

In this phase of the research, the objective was to identify the elements capable of delineating the role of the opinion leaders within the data sets. The reference literature (Cha et al., 2010; Choi, 2015; Leavitt et al., 2009) has not been uniform in its evaluation of the indicators that define opinion leaders. In fact, the follower numbers alone cannot be considered as a value that is able to indicate conclusively the role of the influencer because there are elements that have a certain number of followers exclusively for their social role, so they do not have an interactive audience. It is no coincidence that authors such as Cha (Cha et al., 2010) and Avnit (2009) have talked about the million follower fallacy. Cha (Cha et al., 2010) defines effectiveness in terms of indegree influence, retweet influence, and mention influence. Indegree influence measures the size of the audience for each user, retweet influence indicates the ability of each user to generate viral content, and mention influence implies the ability of every user to be involved in other conversations (Cha et al., 2010).

The indicators defined by Cha (Cha et al., 2010) are reference points for identifying potential opinion leaders in the Twittersphere. However, Lazarsfeld’s general theory (Katz & Lazarsfeld, 1955) defines opinion leaders as the first consumers of media products. Furthermore, as Gladwell (2006) stated, the explosion of a current of thought is determined not only by the law of the few but also by the grip factor and the power of the context.

In relation to these three paradigms (Cha et al., 2010; Gladwell, 2006; Katz & Lazarsfeld, 1955), it was decided also to consider other indicators, creating a synthetic index of “potential opinion leader.” These were the number of post products (media consumption), the number of followers, the number of times posts have been retweeted, and the number of mentions received within the data set (the law of the few). The index is obtained by normalizing the indicators, rescaling them between 0 and 1. The standardization process was realized with the min–max elementary indicators: Z = (x − min(x))/(max(x) − min(x)), and the single elementary indicators were aggregated using the arithmetic mean.

Subsequently, the results of each indicator were cross-referenced with the value of the most retweeted contents (the power of the context), and the common elements present in both data sets (the grip factor) were considered as influencers of the one specific TT and classified in relation to their social role in the digital space of Twitter.

Research Question 4: What are the social roles of opinion leaders in the digital space?

For each opinion leader identified in Phase 3, the specific scope of influence was detected through the qualitative study of their public profile. Moreover, through the analysis of the content of its production in the Twittersphere, the type of opinion expressed in the public sphere of discussion was identified and classified.

Results

The Co-Occurrence of Hashtags

As illustrated in the previous paragraph, the first research step analyzed the hashtags that most frequently co-occurred within the six TTs identified: (1) #Battisti, (2) #Orango, (3) #Rousseau, (4) #Sanremo2019, (5) #Libero, and (6) #SeaWatch3 (see Figure 1). The co-occurrence analysis—using the Gephi software package—illustrates a semantic scene around each network composed of certain hashtags that redefine the topic being discussed, and others that are more general, some of which can be traced back to broader topics of discussion.

OPEN IN VIEWER

Figure 1.

Six semantic graphs, one for each trending topic (#Libero, #Orango, #Battisti, #SeaWatch3, #Sanremo2019, and #Rousseau) visualized by Yifan Hu Proportional layout.

The results also present six dimensions that are not completely separate; thus, some hashtags are common to more than one data set with a preferential position in each network in terms of weighted degree, closeness centrality, and betweenness centrality ¹ (see Table 1). The hashtag #facciamorete is one of the most used in all six TTs. It was created in December 2013 by the digital activist user, Marco Skino (@MPSkino), who in his personal page describes it as a network son of a cultural project aimed at creating a community of commentators who share an anti-fascist philosophy and who are in favor of protecting the Italian Constitution. Another example of transverse hashtags is #salvini, which is the most used in all six data sets. However, it is necessary to point out that Matteo Salvini (@matteosalvini) has a decisive role in four of the six events that led to the birth of the TTs, including, as Secretary of the Lega Party (to which Roberto Calderoli also belongs), the affair related to the hashtag #Orango.

OPEN IN VIEWER

Table 1.

Analysis of the Most Frequent Terms for Each Trending Topic.

Trending Topic	Total Nodes/Edges	Degree Ranking	Weighted Degree Ranking	Closeness Centrality Ranking	Betweenness Centrality Ranking
Battisti	292/3,313
FacciamoRete		5	6	5	5
Salvini		3	3	3	3
SeaWatch3	206/3,186
FacciamoRete		4	2	4	4
Salvini		5	9	4	5
Libero	91/711
FacciamoRete		3	8	3	4
Salvini		12	20	13	10
Orango	32/202
FacciamoRete		3	3	2	3
Salvini		4	6	5	5
Rousseau	382/6,482
FacciamoRete		6	3	6	6
Salvini		4	6	4	4
Sanremo2019	884/14,915
FacciamoRete		21	26	23	24
Salvini		9	7	9	8

To show the transversality of certain elements, an adjacency matrix was built between all the hashtags present within the six TTs, and this was displayed in a graph. From the analysis of the network of all the hashtags present within TT (excluding replication), an indirect small-world graph emerged, composed of 1,735 nodes and 36,933 links, with a diameter of four degrees of separation and with a clustering coefficient of 0.587. The semantic network is composed of semantic links that can be defined as strong links because they are inherent in the same thematic area and in the weaker links of more generic content. Within the graph, eight hashtags have been selected that show a preferential position within the network, both in terms of number of bonds and in terms of weighted degree and betweenness centrality (Table 2).

It should also be noted that two hashtags, #facciamorete and #salvini, appear both as more frequent elements within all six TTs and between nodes that have a preferential position within the hashtag graph. We can, therefore, deduce that #facciamorete and #salvini are two hashtags of relevance, both for structural issues and for the type of content. In fact, their coexistence in several areas of conversation as a frequent element and their placement in a preferential position with respect to the relationship network define these hashtags as elements of influence in the dissemination of information. To the quantitative data is to be added a qualitative reflection concerning the reference public. In fact, #facciamorete and #salvini, by a precise declaration of intent, create two opposite poles of discussion and opinions and are placed as bearers of two specific opinion subculture spheres.

OPEN IN VIEWER

Table 2.

Nodes With a Greater Value of Degree and Betweenness Centrality.

Label	Degree	Weighted Degree	Betweenness Centrality
sanremo2019	1,131	25,903	28,835,222,448
salvini	890	16,069	12,657,211,911
Piattaformarousseau	611	16,009	53,792,735,596
facciamorete	673	12,702	76,748,292,373
rousseau	589	11,720	45,417,590,279

The Relational Structure Inside a Hashtag

The conversational structure of the six TTs was revealed in the relationship between the total number of data of each data set, the number of tweets, and the number of retweets. The frequency of sharing was analyzed for each retweet. From the relationship between the number of tweets and that of retweets, the number of tweets (i.e., the content at “first hand”) had an average of 20.47% compared with 79.53% of retweets (Table 3).

OPEN IN VIEWER

Table 3.

Number of Tweets and Retweets (Internal and Percentage) for Each Trending Topic and the Average Value.

	#Rousseau	#SeaWatch3	#Libero	#Orango	#Sanremo2019	#Battisti	Average
Tweet	20,601	6,903	2,984	898	38,806	10,303	13,415.83
Retweet	64,617	40,305	11,860	8,845	81,341	35,734	40,450.33
Total	85,218	47,208	14,844	9,743	120,147	46,037	53,866.17
Tweet %	24.17	14.62	20.10	9.22	32.30	22.38	20.47
Retweet %	75.83	85.38	79.90	90.78	67.70	77.62	79.53
Total %	100.00	100.00	100.00	100.00	100.00	100.00	100.00

Furthermore, the total number of tweets that had a number of particularly relevant shares represented a very small portion of the whole data set. In fact, on average, the content was retweeted more than 300 times (0.17%); therefore, it reached a wider audience, while 94.57% of the content tended to have a niche visibility not greater than 10 social actors each (Table 4).

OPEN IN VIEWER

Table 4.

Retweet Frequency (Internal and Percentage) for Each Trending Topic and the Average Value.

Frequency	Rousseau	Sea Watch	Libero	Orango	San Remo	Battisti	Average
0–10	19,646	6,397	2,828	834	37,488	9,802	12,832.5
11–100	827	410	127	53	1158	434	501.5
101–200	83	52	12	8	102	42	49.83
201–300	20	21	10	2	27	13	15.5
>300	25	23	7	1	31	12	16.5
Total	20,601	6,903	2,984	898	38,806	10,303	13,415.833
0−10 (%)	95.36	92.67	94.77	92.87	96.60	95.14	94.57
11−100 (%)	40.14	5.94	4.26	5.90	2.98	4.21	4.55
101−200 (%)	0.40	0.75	0.40	0.89	0.26	0.41	0.52
201−300 (%)	0.10	0.30	0.34	0.22	0.07	0.13	0.19
>300 (%)	0.12	0.33	0.23	0.11	0.08	0.12	0.17
Total (%)	100.00	100.00	100.00	100.00	100.00	100.00	100.00

Therefore, it is possible to conclude that despite the breadth of the data sets (323,197 posts), there is little first-hand content compared with the value of the shares. Moreover, few of those contents are able to reach the wider public and, thus, can be considered significant for determining spheres of influence in the Twittersphere. Finally, it is necessary to point out that the values classified as greater than (>) 300, despite being few in number, reach very high sharing values (Table 4).

In view of the highlighted results, it can be stated that the conversational structure of all six TTs represents (Figure 2) a free-scale network ² (Albert & Barabási, 2002). This is composed of a few nodes that have many links representing preferential elements. These have more likelihood of generating greater influence in the network than many nodes with few links, which will have a low level of influence.

OPEN IN VIEWER

Figure 2.

Occurrences curves of the data elaborated in Table 4 in percentage value.

The Opinion Leaders

As previously stated, it is not easy to define useful indicators to identify the figure of the opinion leader within a data set. Indeed, taking the follower number exclusively as a point of reference is completely reductive. Several empirical studies have elaborated synthetic indices using follower, retweet, and mention (Cha et al., 2009; Choi, 2015; Leavitt et al., 2009; Marchetti, cited in Bentivegna, 2014). In the present study, and in accordance with the paradigms of Gladwell (2006) and Katz and Lazarsfeld (1955), it was decided not only to select indicators capable of describing the absolute role of the user but also to evaluate media production, the grip factor, and the power of the context. In other words, it was not only the potential of the users as opinion leaders that was evaluated but also their role as influential elements in a specific context. In this regard, a summary index was drawn up, consisting of the following: the number of followers, the number of posts in the data set, the number of mentions, and the value of the retweets of content relating to the TT (Figure 3).

OPEN IN VIEWER

Figure 3.

A synthesis schema illustrating the process of identifying the opinion leader within the six trending topics.

The index revealed the number of users with the characteristics of the opinion leader as well as the content actually quoted by the public that reached a retweet number greater than 300 (the grip factor). A sample of users was selected from both lists. The analysis carried out highlighted a sample of 26 users, representing the influential elements relating to specific TTs.

In this regard, three observations of an empirical nature can be made. The first concerns the transversality of some actors who are influential in many discussions (i.e., @matteosalvini, @Dio, @lauraboldrini, @ficarraepicone, @francescatotolo, and @vitocontesi). Transversality also encompasses other users (i.e., @makkok, @MPSkino, @Iperbole, and @Manginobrioches) who, despite being in these specific contexts, did not produce content with great potential for virality. The second observation concerns the heterogeneity of the sample; the selected influencers have different social roles: politician, journalist, cartoonist, activist, singer, show character, and comedian. The third observation concerns the users @matteosavinimi and @MPSkino, who, besides being influencers, also represent the central node of the two hashtags #salvini and #Facciamorete, both of which were preferential elements in the hashtag semantic network (Table 5).

OPEN IN VIEWER

Table 5.

Identified Opinion Leaders Referenced by Trending Topic and Their Classification.

Activist	Politician	Journalist	Show Character
@laltrodiego (#Battisti)	@ckyenge (#Orango)	@CucchiRiccardo (#Libero)	@lefrasidiosho (#Rousseau, and #Sanremo2019)
@intuslegens (#SeaWatch3)	@lauraboldrini (#Battisti, #Libero, and #Orango)	@alessioviola (#Libero)	@frafacchinetti (#Sanremo2019)
@francescatotolo (#SeaWatch3 and #Battisti)	@matteosalvinimi (#Battisti, #Sanremo2019, and #SeaWatch3)	@marcellofoa (#SanRemo2019)	@ficarraepicone (#Rousseau and #Libero)
@benq_antonio (#SeaWatch3)	@AlessiaMorani (#Battisti)	@alessiarotta (#Rousseau)
@CaputaLuca (#Libero)		@mattiafeltri (#Battisti)
@SarahConnor_18 (#Rousseau)
@vladiluxuria (#Libero)
@Iperbole (#Libero)
@Manginobrioches (#SeaWatch3)
@Dio (#Rousseau)
@MPSkino (#Rousseau)
@vitocontesi (#SeaWatch3 and #Orango)
@giuliaselvaggi2 (#Orango)
@makkox (#Battisti)

Conclusions

Some of the fundamental elements that regulate the Twittersphere have emerged from this study. The structure of the analyzed discussion networks presents—in its double semantic dimension—the same composition dominated by a few elements with strong relational power. The study, therefore, allows some conclusions to be drawn as to how dialogue is established and maintained in the Twittersphere. The first consideration is the applicability of the Katz and Lazarsfeld (1955) paradigm on the two-stage communication flow and of its use in the digital space (Choi & Park, 2015). Empirical evidence has revealed the presence of molecular social actors who play the role of a sounding board and polarizers of public opinion by generating networks with well-defined contours. Opinion leaders cannot be regarded exclusively as carriers and interpreters of opinions and trends. They are central and aggregating nodes of specific subcultures of opinion and play a predominant role in the management of information and disinformation (Boccia Artieri, 2012).

Having described the real presence and influence of these users, it is necessary to underline the following point: To evaluate the virality of content, one cannot refer solely to the role of the opinion leader; the actual activity of users in a particular area of discussion and the credibility granted to them by the specific audience must be taken into consideration. It is necessary to think not in terms of a receiving public but of receiving publics. Each web medium is the bearer of a precise type of communication, and within it, there are several different audiences. Therefore, the role of the opinion leader is not absolute; it acquires value when it is legitimized by a specific segment of opinion (Gladwell, 2006).

The last but not least consideration concerns the part played by the hashtag in the Twittersphere. According to Bentivegna’s (2015) theory, the hashtag has a dual function: one that is evident and one that is latent. However, if the latent function involves an ad hoc interpretative frame, it is then necessary to add specific niches of discussion that, although defined by the hashtag, are not extemporaneous elements of context but distinctive and transversal to specific affinity networks. Especially in cases such as #salvini and #facciamorete, it is too simplistic to think in terms of information containers or an ad hoc public; rather, they should be visualized as cultural identifiers of certain publics in particular contexts of opinion. Another relevant aspect is that of the structural similarity that exists between the semantic network of hashtags and the structure of the relationship between users. This reflects the model of the invariance scale networks defined by Albert and Barabási (2002), which continues to be the universal, distinctive, and confirmatory element of all digital organisms.

Finally, the research results show that @matteosavinimi and @MPSkino play a particularly influential role in the Twittersphere. The two users appear simultaneously as opinion leaders of some TTs and references of the two transversal hashtags, #salvini and #facciamorete, which in fact describe two opposite opinion polls. For this reason, it is necessary, in future applications, to deepen the role of these two polls as possible permanent communities of opinion in the Twittersphere and reference point to analyze specific collective mass behaviors (Granovetter & Soong, 1983) that arise between the digital space and the offline dimension.