Twitter Users’ Privacy Concerns: What do Their Accounts’ First Names Tell Us?

The decision process for a company must be based on the market preferences, and for this reason, many companies consider that the key to success is to know their customers at a personal level (preferences). A recent study, conducted by IBM, indicates that 48% of consumers want personalized promotions through digital shipments, while 44% expect personalized benefits in physical stores (Gonzalez, 2015).

One critical strategy to renew and stay in the commercial race in today’s competitive market is to know how the market is evolving. Customers provoke the evolution of the market, and for this reason companies are investing large amounts of money in business intelligent tools that allow them to predict the preferences of their customers. If companies know the preferences of their customers, they can anticipate changes, assess risks, reduce costs, and increase profits.

According to Howland, CEOs must understand future consumers’ needs in order to anticipate trends in the market (Howland, 2014). Companies must invest time in learning about consumer needs, wants, and demands. If they know this, companies will have a “consistent service through strategic dialogue” (Howland, 2014). This will increase consumer loyalty and will attract new potential consumers.

The magazine Entrepreneur remarks that market segmentation is the best strategy to gain an understanding of the customers’ preferences (Entrepreneur Magazine, 2012). Fran Leo defines market segmentation as the process of dividing the total market of a good or service into several smaller, internally homogeneous groups such as by customers’ ages, genders, and lifestyles (Ale, 2015). One of the main objectives of market segmentation is to figure out the preferences of the users in order to provide a more personalized service without the need of conducting direct dialogues. If the segmentation is made properly, the company will reduce the cost of targeted advertising, offer more personalized service, and potentially increase its revenues. For example, previous work has shown that gender plays an important role in the purchase decision-making. Van Aswegen explains that shopping behavior in men and women is highly affected by their considerations, concerns, perspectives, and motives. When companies are developing target ads strategies must recognize gender-specific concerns, to offer what the users really want (van Aswegen, 2015).

On the other hand, today’s global market makes it challenging for companies to have direct communication with each customer to know his or her personal preferences, and allocate him/her into one of the identified homogenous groups. For this reason, companies are using Internet users’ social media data that allow them to know their customers, without using a form of direct communication and offer them a more personalized service. However, companies face a big challenge in this approach, that is, the Internet users’ privacy concern.

For Parker, there is not consensus in the legal and philosophical literature about what is the proper definition of personal privacy (Parker, 1973). Gross and Acquisti said that there are various levels and aspects regarding privacy on social platforms (Gross & Acquisti, 2005). In the social media context, privacy can be understood as the right to control how personal information is collected and used by social platforms.

Although many personal data are freely available on the Internet and easily accessible, this does not mean that the users are not concerned about their privacy. For Greenfield, it is important that companies “don’t be creepy using robust user data for ad targeting while respecting privacy” (Greenfield, 2012). Greenfield mentions that there is a fine line between creepy and appropriate. In the online environment, some users are more likely to receive targeted ads and some are not. Greenfield recommends companies to respect the users who are not likely to receive targeted ads because this can cause to lose a customer by a simple ad (Greenfield, 2012). Morey, Forbath, and Schoop (2015) explain the importance of companies to be trusted by their consumers. The authors point out that consumers expressed concern about their privacy. American users rank privacy issues high; 72% prefer not to share personal information with business (Morey, Forbath, & Schoop, 2015). It is important for companies to know their customers in a personal level in order to offer personalized services, but how companies can offer target ads if certain users are not likely to share their personal information. How to identify these users from the pool of Internet users then becomes an important and urgent research question in this direction.

Interested in answering this question, it has been developing computational techniques to help classify Twitter public accounts into privacy-concerned and privacy-not-concerned groups. Twitter has over 330 million active users (Statista, 2016). When an individual creates an account on Twitter, they only must type their complete name, email address, and password.

In 2014, Twitter announced, “advertisers can use Twitter Analytics data to target users more directly and reach the most valuable audiences” (Lopez, 2014). While Twitter users can choose to make their accounts protected, those with public setting can still be privacy concerned—e.g., research has shown that privacy setting in social media is not an accurate indicator of the user’s level of privacy concern (Gross & Acquisti, 2005b; Irani et al., 2009).

Compared to other social networks, Twitter works with basic privacy setting: protected or public account. When a user creates an account, his/her information is publicly accessible by default. It is necessary to change the profile settings to make a profile private. Michael Zimmer and Nicholas Proferes argue that not all the users are aware about this default option of privacy, and accidentally can provide more personal information than they intended (Zimmer & Proferes, 2014). For the intentions of this work, we considered users whose profiles are private as users with high levels of privacy concern.

Consumers who go beyond to protect their information in social platforms are not happy about companies knowing their personal information (Lopez, 2014). These consumers’ attitudes reveal concerns regarding the privacy use of their data. Prior work has leveraged features from Twitter’s tweets and some profile attributes to identify a Twitter public account’s level of privacy concern (Khazaei et al. 2016b).

In this paper, we examine whether the first names and profile descriptions in Twitter accounts as clues to infer the user gender provide useful information for predicting the privacy settings of the user. Our work builds on the previous studies that demonstrate the correlation between a person’s gender and the level of his/her privacy concern. Many studies indicate that women tend to be more concerned about online privacy and security issues than men (Adam, 2000; Sieger & Moller, 2012). But when the sites are related to shopping or fashion women-related online behaviors are contradictory with their alleged high privacy concerns (Katell, Mishra, & Scaff, 2016; Nazir et al., 2012). Based on 398 online consumer interviews, Riquelme and Roman found that the influence of both privacy and security on online trust was stronger for younger, more educated, and less extroverted males (Riquelme & Román, 2014).

In our work, we use the first name and profile description in a Twitter user account as the main clues to infer the gender of the Twitter user. This approach, that is, to infer the gender in Twitter users through the analysis of the first name, has a reported accuracy of 80.48% (Liu & Ruths, 2013). Previous works have shown that feature combinations of profile attributes increase the accuracy of the result of gender inference (Fernandez, Moctezuma, & Sordia, 2016). We analyzed the profile descriptions through counting the number of feminine and masculine words, and using a Bayesian network with weighted word frequencies to infer the gender of a user. With these results (inferred gender by name, and by profile description), we estimate the gender of a user using Neural Networks algorithm and obtaining 83.47% of accuracy.

Then, we analyzed the correlation between the inferred gender of a user and his/her level of privacy concern. In addition, with the resulting set of features, we create a model to predict user privacy settings. Our best classifier obtained an F-score of 0.73, improving in 0.002 points the accuracy of previous classifiers (Khazaei et al., 2016b). This paper is organized as follows. In Section 2, we present some of the related works that are the starting point of our research. In Section 3, we describe our data sets, and process for labeling these data. In Section 4, we explain our methodology for gender recognition. In Section 5, we report our results related to gender and privacy concerns. We summarize our findings and conclude in Section 6. Finally, Section 7 deals with future directions of our work, especially to the accuracy of our method and give detailed advice to companies to help them improve their strategies for market segmentation.

After constructing the name database, we wrote a Java program to infer the gender of each user in our Twitter user database. We connected our program to the name database to analyze the user names and infer their gender. The program first extracts the complete name of each user and divides it in sub. Then it takes the first substring as the first name and compares it with the names in our name database. It returns the gender if it is found in the name database and tags the user with the gender. If the name is not found in the database, then the user is tagged as “undefined”.

Because a name is not enough to infer the gender of a user, we analyzed the profile description. Based on the findings of (Argamon et al., 2006), we wrote a Java program that uses a Bayesian Network with weighted word frequencies and part of speech to estimate the gender of an author. While most of the profile descriptions were in English or Spanish, four users’ descriptions were not. We removed them from the data set. Inspired by Gaucher, Friesen, and Kay (2011)’s work, we built our own list of masculine and female words for English and Spanish. It was built based on gender stereotypes and social roles. The female words list contains words such as “wife”, “mother”, “daughter”, “girl”, “lovely”, “delicate”, etc. The male words list contains words such as “husband”, “brother”, “son”, “father”, “boy”, “player”, etc. Our English lists have a total of 67 masculine words, and 75 feminine words. The Spanish version of our list has less number of words because for this language the ending of an adjective determines its gender. We then counted the frequency of masculine and feminine words in the profile description depending on its language and the corresponding gender word list.

After analyzing the name and the profile description, we added two new columns to our three data sets. These columns represented the results of our gender analysis. The training data set (data set C) has three columns that reflect the gender of a user: gender by the human annotators, gender by name search, and gender by profile description analysis. We developed classifiers using this subset of features to obtain the highest accuracy possible regarding to the gender estimation. Table 1 shows the results of the four classifiers. As is shown in the table, Neural Networks classifier had the best performance. We used it to classify the users of our other data sets.

Table 1

Evaluation of classification results.

We next explored whether and how inferred gender is linked to the privacy settings with data set A. Our first null hypothesis was that the gender of a user, which is reflected from his/her first name and profile description, is independent of the account’s privacy setting. We tested our null hypothesis using the chi-square test.

Using the data set A, we analyzed the users whose names do not reveal a specific gender. We were interested in examining whether the fact that a name that does not exist in the name dictionary suggests some level of user awareness of privacy. The data set has one field called HasARealName which represents whether or not any part of the account name can be found in a directory of English names (Khazaei et al., 2016b). The value of this field is one if any part of the account name does appear in the name directory, or zero if it does not. Our second null hypothesis was that whether or not the name exists in the name dictionary is independent of the user’s privacy setting.

Lastly, we explored the correlation between a user account’s privacy setting and the inferred gender, and then added the inferred gender feature to Khazaei et al. (2016a)’s privacy preference prediction model.

Our results of the hypotheses testing and the performance of the revised model are presented in the results section.

At a 0.005 significance level, our chi-square test shows that the gender of a user, reflected from his/her first name and the profile description, is correlated with the user account’s privacy setting (P < 0.0001,χ² = 14075.06). The Cramer’s V measure of our chi-square test’s effect size is 0.0089 (DF = 3), which is considered to be a small effect (Cohen, 1988). As shown in Table 2, female users tend to be protected users. This is consistent with the findings from the related literature. For example, Adam (2000) states that women are more aware of their personal protection because they consider privacy issues more than men do.

Table 2

User accounts’ privacy setting (protected vs public) and their gender inferred from their names and profile descriptions.

Our examination of the undefined users shows that they tend to use names that do not exist in the name dictionary and have protected accounts. Table 3 shows undefined protected users against undefined public users and the use of a name that exists or non-exists. At 0.005 significance level, we rejected the second null hypothesis (p < 0.0001). This result indicates that users whose gender cannot be inferred from the names tend to be aware of the privacy setting.

Table 3

Users whose genders cannot be inferred from the names and the privacy setting of their accounts.

In the last task of identifying the predictive power of the inferred gender for the account’s privacy setting, we considered two categories for the inferred gender: RevealedGender that includes users identified as females or males, and or NotRevealedGender that includes only the users tagged as undefined. Our Spearman correlation test showed that there is a negative correlation between the inferred gender and the account’s privacy ratio (r = -0.113, p < 0.0005), which means that the higher the privacy ratio (i.e., the more privacy concerned) the less likely that the gender can be inferred from the names and the profile descriptions.

Khazaei et al. (2016a) proposed a model that used data set A to predict a Twitter user’s privacy setting based on the privacy ratios of various profile attributes and on the linguistic features of the profile descriptions. With this data set, we added gender as an additional feature and examined the added predictive power. The comparison of the results is provided in Table 4. As is shown in the table, the addition of the inferred gender feature improved the prediction performance slightly by about 2%.

Table 4

Comparison of our classification results with Khazaei et al. (2016a).

It is known that not all twitter accounts are real user accounts, and social bots tweet as well (Chu et al., 2010). A major limitation of our study is the lack of consideration of social bots in the data. In our study, this implies that at least some percentage of the undefined accounts, that is, accounts that had names non-existent in the name dictionary, are social bots. It will be interesting to explore the privacy setting of social bots in the Twitter space.

Nonetheless, our study suggests that companies that are conscientious of the various levels of customers’ privacy concerns may leverage the provided names and profile descriptions in Twitter accounts to infer the accounts’ privacy concern.