Privacy concerns in social media UGC communities: Understanding user behavior sentiments in complex networks

The data extraction and analysis process has followed the procedure previously proposed by Hiremath and Patil (2020). Specifically, we connected to the public Twitter API to download pieces of UGC in the form of tweets with the hashtag #Privacy #PrivacyConcerns. The queries containing the indicated hashtags were made from May 5 to May 15, 2022. These dates were used because there was no international event linked to user privacy or similar issues that could have changed the results (Kim et al. 2013).

The terms used to collect the data from Twitter matched our research questions and objectives (Saura et al. 2022b). For the development of the data collection, a non-probability sampling frame was followed (Schillewaert et al. 1998). As indicated in Saura et al. (2022c), “non-probability sampling is an approach to sample elaboration, which is also known as simple judgment” (p. 3). In this study, we followed the criteria formulated by Lehdonvirta et al. (2021) and Saura et al. (2022c) to develop the sampling process in which we considered non-probabilistic variables. Lehdonvirta et al. (2021) indicated that non-probability sampling is commonly used in studies that work with UGC from social networks.

The data collection process was developed with Mac version of Python 3.7.0. Once the data were collected, the tweets were filtered. First, images and videos were excluded, since this study focused on NPL, thus not covering visual or interactive elements of the database. This is a common process in studies that analyze Twitter data (Rafail 2018). In addition, links and URLs were removed from the database in order to clean the content. Additional filtering processes were applied using Python and Pandas libraries. Only those tweets that belonged to UGC online communities on Twitter were analyzed, taking into account the ties and connections existing in the users and the content generated around the community.

Similarly, repeated tweets were eliminated, as were also emojis to enhance the quality of the data. In this way, from the database of 75,307 tweets, after the filtering process and identification of communities, a total UGC sample of 49,701 tweets was obtained.

To apply data visualization algorithms and online communities, we used the algorithm developed by Blondel et al. (2008). This algorithm makes it possible to identify communities in complex networks through the modularity report (MR) indicator. This technique uses a resolution enhancer to improve visualization of the algorithm based on Lambiotte et al. (2014). This type of algorithm used to visually represent online communities was previously used by Jacomy et al. (2014).

In the present study, the main objective of the identification of online communities was to understand the role played by privacy concerns of users and their online behavior. In this way, our sample was structured into nodes related to each other. Nodes are neurons that represent links between different users and which, when analyzed together, form communities on Twitter (Tidke et al. 2020) (see Fig. 1a, b). Figure 1a shows the global community of users linked by the nodes according to their connections. Figure 1b shows these in an enlarged way, thus identifying sub-communities that make up the global community and that determine topics of conversations, interests, and comments.

The link between these nodes and the relevance and weight they have over the remaining nodes is what determines their relationships in terms of relevance in the total database (Java et al. 2007). Accordingly, the nodes indicate the communities and the themes that compose them. UGC communities in social media can be visually represented both by the number of users who make up a node and the number of topics that make up the set of nodes in an online community (Barbosa et al. 2022; Guerola-Navarro and Stratu-Strelet 2022).

As concerns the resolution of the algorithm and its application, it tries to identify the sub-communities based on the weight of nodes linked to the main community, which, in the present study, is that of online privacy concerns. According to Ausserhofer and Maireder (2013), the higher the visualization and its resolution when the algorithm is applied, the fewer but more specific communities are obtained (Gu et al. 2022). Conversely, the lower the resolution, the more communities, as well as the more users and their links are identified by the algorithm (Chawra and Gupta 2022).

As demonstrated by Fogel and Nehmad (2009), networks of nodes can be visually understood as systematically organized communities around the union of nodes. The organizational utility of this type of system is that it analyzes the links between each node and their distance in terms of weight (Trott 2022). One node can be identified in relation to several descriptors, which makes it possible to demonstrate the themes around which the communities are grouped, as well as their classification. The richer in terms of data a community is, the more connections are there between its nodes (Mishra et al. 2011; Lam and Chow 2022).

As noted by Tokarchuk et al. (2022), there are different approaches to divide a database into different sentiments. In the present study, we used the theoretical framework of NLP and CATA, which is a sentiment analysis of text (i.e., excluding visual and multimedia elements in a network). Following Hiremath and Patil (2020) and Saura et al. (2022b), sentiment analysis was performed using the Textblob algorithm (Rasmusen et al. 2022; John and Sam 2022). This algorithm, developed in Python with NLTK and patterns, has been widely used in the literature.

However, a limitation of sentiment analysis is that it fails to identify connotations or sarcasm, as well as irony (Hussein 2018). The Textblob algorithm was trained, and these issues were taken into account in order to proceed with an efficient and accurate development of the results. In order to develop the algorithm with high precision, the tweets were rated based on their polarity and subjectivity. Polarity was rated from − 1 to 1 and subjectivity from 0 to 1.0 (Hiremath and Patil 2020; Saura et al. 2022b).

The algorithm was trained a total of 935 times using tweets that had been manually classified (Neethu and Rajasree 2013). That is, manually classified tweets were used as inputs. Upon learning, the algorithm classified the remaining tweets into three groups depending on the sentiment expressed in them (positive, negative, and neutral). As argued by Taboada (2016), feelings can be extended to other emotions or moods. However, following Hiremath and Patil (2020), in the present study, we decided to focus on the analysis of only three aforementioned feelings. Similarly, aiming to increase precision of our sentiment analysis development process, five cross-validation experiments were performed (Saura et al. 2022b; Mohammed et al. 2022).

Specifically, precision, recall, f-1 score, and support were used as measurement factors and variables. Sentiment analysis results were obtained in terms of macro average and weighted average. The models used as experiments in the 5 cross validation experiments were as follows: support vector classifier (Sarkar et al. 2019), multinomial naïve Bayes (Griol et al. 2017), logistic regression (Sarkar et al. 2019), and random forest classifier (Vijayarani and Janani 2016).

Using the approach proposed by Blondel et al. (2008) to the sample, we identified user communities with different resolutions. The objective was to configure the largest number of relevant user communities found in the sample in relation to privacy and user behavior. In order to obtain the best results, four different tests were performed applying the algorithm in different resolutions. The results were interpreted based on the number of communities, modularity report, modularity resolution, maximum modularity class, and minimum modularity class (Bouarara 2021).

The modularity report indicates the relevance and weight of the identified community: the higher the modularity, the higher the relevance of a community in the database. Modularity resolution indicates the resolution of the algorithm used to visualize the communities. Maximum and minimum modularity class indicates the maximum and minimum relevance value identified in the total of the identified communities (Blondel et al. 2008). The results of five tests that were performed are shown in Table 1.

Aiming to improve the visualization of the neural network and the identified communities, several filters were implemented so as to enlarge the resolution of the identified communities using Gephi 0.9.7. Following Jacomy et al. (2014), in order to modify the variable modularity report and the algorithm in charge of visualizing the data, we used Force Atlas and Force Atlas2. These filters can be applied with Gephi software. The following data visualization filters were applied: threads number (1), tolerance (1.0), approximation (0.1), scaling (0.2), gravity (1.0), prevent overlap (True) and edge weight influence (4.0). The size of the resolution in relation to the visualization of each node was 10 points. The number of nodes analyzed were 9860 and 141,551 edges. Dynamic graph and multi graph attributes were not applied. The type of graph was undirected using auto-scale. The edges merge strategy was sum. The total time consumed to apply Force Atlas and Force Atlas2 algorithm before they were stopped were 3 min. The nodes were then grouped into communities in relation to their weight in the identified community and were displayed in different colors. In terms of weight in the MR, the communities were organized as follows: privacy concerns (MR = 140), privacy settings (MR = 136), personal information (MR = 116), hacking (MR = 113), social engineering (MR = 110), protection software (MR = 107), false information (MR = 101), cyber-bulling (MR = 97), impersonation (MR = 96), cookies data (MR = 96), and eCommerce (MR = 94). In is important to notice that the Fig. 2 has been graphically adapted to show the identified communities accordantly. In Fig. 2, the identified communities are shown. The size of each community is represented by the number of nodes that compose it. The distance between each community as well as the distance to the center of the neural network, determine the relationship and the proximity of the topic to privacy concerns.

Next, we performed sentiment analysis. The accuracy variable was identified as the metric that verifies the quality of the sentiment analysis process (Sharma et al. 2021). In the present study, the accuracy variable was used to measure the effectiveness of algorithms that work with machine learning. The five technologies used to validate the analysis are presented in Table 2 with Sl. No. from 0 to 4. The table shows the number of the experiments, the name of the model used, the fold_idx value that represents the order of the text and, finally, the accuracy obtained in Textblob (Alowibdi et al. 2021).

In line with Hiremath and Patil (2020) and Saura et al. (2022b), the highest accuracy value obtained was in relation to linear SVC Sl. No.7 (0.880057). For RFC, the results were Sl. No.2 (0.790031). As concerns LG, the precision of 0.800979 was obtained in Experiment 3. Finally, with regard to multinomial naïve Bayes, the highest result was Sl. No. 4, with the accuracy value of 0.800979.

To summarize and facilitate the understanding of the results of the sentiment analysis, Table 4 presents the names of the models used to develop the sentiment analysis with Textblob and the highest scores obtained for accuracy in each case. As can be seen in Table 3, the accuracy of LinearSVC was 0.880057 the highest and was the used for the development of the study.

Furthermore, it is a standard procedure in research carried out with NLP and CATA to represent the variables of recall, f1-score, and support (along with accuracy). In the present study, the sentiment analysis algorithm was configured around three sentiments (positive, negative and neutral). The indicated variables were therefore linked to each of these feelings (Chen et al. 2022). As argued by Lease (2011), the accuracy variable refers to the quality of the machine-learning algorithm, while the recall variable measures the amount of machine learning (such as technology) used to support the development of the model applied to the database of tweets.

Furthermore, the f1-score variable is a metric that combines both the accuracy and recall variables in a single analytical indicator (Heikal and Eldawlatly 2020; Singh and Sachan 2021). Therefore, using f1-score makes it possible to compare different behaviors of the algorithm in relation to the accuracy of the model used. Similarly, the support variable determines how much machine learning is used by the model to make the predictions with which the accuracy is computed. Table 4 presents these variables, as well as the weighted average measures associated with relativity in terms of weight.

In order to be able to divide the identified communities into sentiments, all tweets in the database linked to each of the identified communities were selected and divided into different databases. In this way, once the tweets of the communities in different databases were extracted, the sentiment analysis was applied to the set of tweets contained in these databases. Table 5 lists the identified communities, their descriptions, and the associated sentiments.

The results of the present study can be used by other researchers to explore new challenges and opportunities in the industry of privacy in digital environments and user behavior. Specifically, in future studies, the identified communities can be used as variables and/or constructs of empirical and quantitative models that could be empirically tested. In addition, the communities identified through the algorithm of complex social networks ForceAtlas2 allow other researchers to replicate these techniques in future studies.

Finally, the sentiments associated with the identified communities identified and the weight of each community can help other researchers to set more specific research objectives to address opportunities and issues related to privacy in a more specific and detailed way.

As concerns practical implications, this study provides valuable and practical insights into some techniques and challenges in digital environments related to users’ personal data. Similarly, by understanding the communities involved, general users and practitioners, as well as digital agencies and public institutions linked to digital business models, can better understand how users share their opinions on the Internet and organize themselves specifically in Twitter.

Accordingly, our results can help both companies to develop their strategies and individuals to form and develop safer browsing habits. Likewise, the results of this study can be used to improve the security and communication protocols of public institutions that are dedicated to the education of healthy navigation habits. Institutions such as the United Nations or the European Commission in Europe can use the results identified to increase investment in programs to educate on security on social network.

Limitation of the present study are related to the number of tweets in our sample, as well as to the application and filtering according to the resolution used in the neural network identification algorithm used in Gephi. In the future, in order to increase precision of the results, the same research design could be applied to a larger sample and improved algorithms based on machine-learning technology.