WhatsApp usage patterns and prediction of demographic characteristics without access to message content

Abstract

BACKGROUND Social networks on the Internet have become ubiquitous applications that allow people to easily share text, pictures, and audio and video files. Popular networks include WhatsApp, Facebook, Reddit, and LinkedIn. OBJECTIVE We present an extensive study of the usage of the WhatsApp social network, an Internet messaging application that is quickly replacing SMS (short message service) messaging. To better understand people's use of the network, we provide an analysis of over 6 million encrypted messages from over 100 users, with the objective of building demographic prediction models that use activity data but not the content of these messages. METHODS We performed extensive statistical and numerical analysis of the data and found significant differences in WhatsApp usage across people of different genders and ages. We also entered the data into the Weka and pROC data mining packages and studied models created from decision trees, Bayesian networks, and logistic regression algorithms. RESULTS We found that different gender and age demographics had significantly different usage habits in almost all message and group attributes. We also noted differences in users' group behavior and created prediction models, including the likelihood that a given group would have relatively more file attachments and if a group would contain a larger number of participants, a higher frequency of activity, quicker response times, and shorter messages. ONCLUSIONS We were successful in quantifying and predicting a user's gender and age demographic. Similarly, we were able to predict different types of group usage. All models were built without analyzing message content. CONTRIBUTION The main contribution of this paper is the ability to predict user demographics without having access to users' text content. We present a detailed discussion about the specific attributes that were contained in all predictive models and suggest possible applications based on these results.

Headnote

Abstract

BACKGROUND

Social networks on the Internet have become ubiquitous applications that allow people to easily share text, pictures, and audio and video files. Popular networks include WhatsApp, Facebook, Reddit, and LinkedIn.

OBJECTIVE

We present an extensive study of the usage of the WhatsApp social network, an Internet messaging application that is quickly replacing SMS (short message service) messaging. To better understand people's use of the network, we provide an analysis of over 6 million encrypted messages from over 100 users, with the objective of building demographic prediction models that use activity data but not the content of these messages.

METHODS

We performed extensive statistical and numerical analysis of the data and found significant differences in WhatsApp usage across people of different genders and ages. We also entered the data into the Weka and pROC data mining packages and studied models created from decision trees, Bayesian networks, and logistic regression algorithms.

RESULTS

We found that different gender and age demographics had significantly different usage habits in almost all message and group attributes. We also noted differences in users' group behavior and created prediction models, including the likelihood that a given group would have relatively more file attachments and if a group would contain a larger number of participants, a higher frequency of activity, quicker response times, and shorter messages.

CONCLUSIONS

We were successful in quantifying and predicting a user's gender and age demographic. Similarly, we were able to predict different types of group usage. All models were built without analyzing message content.

CONTRIBUTION

The main contribution of this paper is the ability to predict user demographics without having access to users' text content. We present a detailed discussion about the specific attributes that were contained in all predictive models and suggest possible applications based on these results.

1.Introduction

Social networks on the Internet have become ubiquitous applications, allowing people to easily share text, pictures, and audio and video files. Popular networks include Facebook, Reddit, and LinkedIn, all of which maintain websites that serve as hubs, facilitating people's information sharing. In contrast, the relatively new WhatsApp application is a smartphone application that enables people to share information directly via their phones. Since its introduction in 2009, its growth has steadily increased, and as of April 2016, it has a user base of over a billion monthly active users.3 Although many alternatives to WhatsApp are currently available in different online application stores (e.g., Kik, Telegram, Line Messenger, BBM, WeChat), WhatsApp is currently the most popular messaging application with the largest name recognition, by far the largest user base, and the strongest corporate backing since its acquisition by Facebook in 2014. Given the emerging importance of this network, it is not surprising that there is a growing interest in researching it, including user studies about people's WhatsApp use and possible applications (Jain, Eddy Luaran, and Rahman 2016; Gulacti et al. 2016; Fiadino, Schiavone, and Casas 2014; Church and de Oliveira 2013; Pielot et al. 2014; O'Hara et al. 2014; Bouhnik and Deshen 2014; Mudliar and Rangaswamy 2015; Montag et al. 2015; Johnston et al. 2015).

This paper's main contribution is that we have successfully created models that predict WhatsApp usage patterns between different types of users and groups without relying on the content of people's text messages. As discussed in more detail in the following section, prior studies about WhatsApp typically based their analysis on the content within the messages (Wang, Burke, and Kraut 2013; Argamon et al. 2009; Wagner et al. 2015). Collecting and storing text messages is problematic for several reasons. First, privacy concerns exist in storing and analyzing people's messages and can raise significant ethical concerns (Van Wel and Royakkers 2004). Second, storing all information from peoples' text messages can require large amounts of storage space, which in turn increases the cost of such analyses (Fan et al. 2006). Instead, we focus exclusively on general message information, such as the message's length, the size of the conversation group to which it was sent, and temporal properties such as the time it was sent and how much time elapsed between a given message and the previous one. Despite the lack of content, we successfully created models that predict usage patterns for different types of users and groups. In previous studies, such patterns were found by checking a specific thesis via distributing and analyzing targeted questions from questionnaires (Church and de Oliveira 2013; Pielot et al. 2014; O'Hara et al. 2014; Mudliar and Rangaswamy 2015). These previously used methods are significantly more time-intensive than the automated machine learning approach that we used. While this methodology has been used to study other social networks, including Facebook (Wang, Burke, and Kraut 2013; Xiang, Neville, and Rogati 2010; Bakshy et al. 2012) and MySpace (Thelwall, Wilkinson, and Uppal 2010), applying the methodology to the WhatsApp network is significantly more complicated because, in contrast to these other networks, no public dataset currently exists, probably because of the medium involved. Whereas other social networks are primarily web-based, enabling data to be compiled through web crawling, the WhatsApp network is based on individuals' private phone use and thus is not publicly available. Furthermore, these studies typically use the messages' actual content, something we intentionally did not use.

As we further describe in the following sections, we performed an in-depth study based on WhatsApp messages and conversation groups by collecting over 6 million WhatsApp messages from 111 students between the ages of 18 and 34. All messages were encrypted with the HMAC hash function, making it impossible to discern the messages' content. Even without using the messages' content, our analysis revealed several key insights. First, we did in fact find significant differences in WhatsApp usage profiles across people of different genders and ages. Second, we generated predictive models for different types of WhatsApp usage to demonstrate that these types of models could be built by applying machine learning and data mining tools on WhatsApp data when collected at the message level. Specifically, we entered the data into the Weka data mining package (Witten and Frank 2005) and studied the output from decision tree and Bayesian network algorithms. Additionally, we generated logistic regression models using the pROC package in R (Robin et al. 2011). Despite our lack of relying on any user-generated content whatsoever, these algorithms were successful in building models that can accurately predict a person's gender and approximate age. They were also successful in predicting which WhatsApp groups have certain qualities, such as higher percentages of file attachments, quicker responses, larger discussion groups, and shorter messages. One key advantage in analyzing the results from the decision tree algorithm is that it generates an assessment of which attributes and logical rules were important in building these prediction models, thereby providing additional insights. Last, we note the importance of these results with possible future directions and applications.

2.Background

This work's primary focus is to address how individuals behave on WhatsApp without using the content of their messages. The WhatsApp social network is unique in several ways. WhatsApp was developed to allow users to privately and freely send messages to each other through their smartphones. It provides a free alternative to SMS (short message services) which is often still a metered (pay per use) service. Not only is WhatsApp often more cost effective than SMS, but it facilitates large group conversations, something that is difficult through SMS, if not impossible. While freely sharing information over the Internet is common to many social networks, and other public messaging services (such as Twitter) exist, the private nature of the WhatsApp network makes it rather unique. Another difference between WhatsApp and other social networks is that membership is created and updated directly via people's smartphones. Not only is registration done exclusively through one's phone number, but the smartphone is the primary interface for sending and receiving messages.4 Also, WhatsApp interpersonal conversation groups are the network's only communication medium. Groups are formed by adding people's telephone numbers to that group. In contrast, other social networks are based on user membership and primarily focus on public messages (called 'posts' on Facebook and 'tweets' on Twitter), where information is sent to all connected rather than through private groups. Furthermore, Facebook is a network for publicly sharing photos, updates, and general news with members who "follow" you. Twitter is a microblog network where members interact through concise messages of up to 280 characters. Given these and other differences between WhatsApp and other social networks, we believe that existing research about other networks is not necessarily applicable, and a new and thorough analysis of WhatsApp is warranted.

Much recent work has been dedicated to the study of how people use WhatsApp and the role of this new application in social communication. Most works to date have analyzed peoples' behavior through conducting surveys and targeted interviews. For example, work by Church and de Oliveira (2013) conducted an online survey asking users targeted questions that were aimed at understanding differences between WhatsApp and SMS usage. Pielot et al. (2014) created a survey focusing on the question of whether people expected an answer to their WhatsApp and SMS messages within several minutes. O'Hara et al. (2014) interviewed 20 WhatsApp users for nearly an hour each, asking them semi-structured questions aimed at determining the nature of relationships forged with the people with whom they communicated. Mudliar and Rangaswamy (2015) spent over 350 hours observing 109 students and conducted surveys to understand gender differences within Indian students' use of WhatsApp. All of these studies can be characterized as being formed to answer specific questions by conducting targeted surveys and interviews.

Our work is unique in that it uses statistical and data mining methods to study WhatsApp usage at the message level without knowing the content of the messages. Motivation for our study, as for previous WhatsApp research, is to analyze differences between genders, the time that elapses until a message is answered, and the characteristics of larger and smaller discussion groups. However, our study is fundamentally different in that it is based solely on actual WhatsApp metamessage data, and creates predictive models without any content knowledge. The issue of smartphone usage analysis was recently studied, and one of the study's conclusions was that people often inaccurately report their own usage in questionnaires (Lin et al. 2015). Our methodology helps avoid this issue. To our knowledge, only one other study, performed by Montag et al. (2015), logged WhatsApp usage from nearly 2,500 participants. While the number of participants in this study is impressive, the actual data logged was significantly less robust than in this study. They collected only general metadata about use, limited information about WhatsApp messages and no information about the users' group activity.

In theory, even more accurate models could have been constructed had we also analyzed the messages' content. Specifically, models that were previously developed can predict a user's gender, age, native language, or personality (Wang, Burke, and Kraut 2013; Argamon et al. 2009) based on content. Examples include work by Argamon et al. (2009), which focused on creating models that identify word usage differences between men and women on Internet blogs. Similarly, Wagner et al. (2015) focused on content differences between men and women in Wikipedia, and Wang, Burke, and Kraut performed a study of content differences between genders on Facebook (Wang, Burke, and Kraut 2013). However, as the WhatsApp network is inherently private, such approaches could not be applied in our case due to privacy concerns. As we now detail, even without this information we were successful overall in predicting a user's demographic and group behavior.

3.Dataset creation and description

Given the private nature of the WhatsApp network, this study's first challenge was to create a WhatsApp message dataset while still ensuring users' privacy. To do so, we developed software that integrated with the Android Debug Bridge (ADB), which is an external tool that can back up an Android application.5 This enabled taking a "snapshot" of a person's groups and messages as they appear in that person's phone. To make the data anonymous, the software encrypts the data that was pulled directly from the participant's smartphone by using the hash-based message authentication code (HMAC). The entire process of obtaining a participant's data lasted approximately 15 minutes and we compensated each participant $12 for their time and temporary inability to use their phones. We also collected the participants' general demographic information including their age, gender, place of residence, and educational background. In addition, we asked them to self-rate their sociability and WhatsApp usage on a five-point Likert scale (Low to High), and to answer four Boolean questions dealing with whether they use WhatsApp for communication with work, family, friends, or others. To guarantee that the study complied with policies on ethical conduct, we obtained institutional review board (IRB) approval before beginning data collection.

We found it challenging to recruit participants because people were quite reluctant to provide information about their WhatsApp messages, even when we emphasized that all content sent was encrypted, and that no nonencrypted content data was ever sent. While we attempted to recruit participants from all age groups, we found that student participants, found through advertisements on campus, were the demographic most willing to participate. Nonetheless, we did make a concerted effort to find people in other demographics through word of mouth. Through this process we recruited a total of 137 participants. Only 19 of these participants were not college-age students (18 through 34), so we removed these participants' data from the analysis because this group was not large enough to be validly divided into further age subgroups. Thus, we are aware that the data collection process was biased for younger people, and we hope to address this in the future through a different collection process for other age groups.

To remove any biases in our analysis from people who had not used WhatsApp for long periods of time or who did not generally engage in WhatsApp conversations, we further removed another 7 people who were active in WhatsApp for under 20 days or had fewer than ten total WhatsApp groups. Thus, the dataset in this study contains messages from 111 participants, of which 59 were female and 52 were male, all of them young adults between 18 and 34 years of age, with a median age of 27. The 111 participants sent and received a total of 6,449,631 messages over an average period of approximately 15 months.6

The defining characteristic for the logged data is that it intentionally contains no textual content. All types of textual content are unavailable, including any special characters or emojis that exist in the messages. Similarly, we stress that we have no information about the message recipients other than an anonymous ID because all data is anonymous.

While we did not have the messages' content or recipient information, we were nonetheless still able to glean a great deal of usage information regarding message and group statistics. The first type of information focused on general information surrounding the messages' characteristics, such as when they were sent, the number of words in the message, whether the message included a file, and the length of time that elapsed before a response was sent. Once we had all of the messages, we discretized their time into categories based on the percentages of messages sent over each hour-long interval (e.g., messages sent between 5:00pm and 6:00pm). Similarly, we discretized the number of messages into the categories of 1, 2, 3-5, 6-10, 11-20, and 20+ words. We then discretized messages according to the time that elapsed between messages - under 1, 1-2, 3-5, 6-15, 16-30, and 31-60 minutes. The motivation behind this process is our assumption that messages that appear within a relatively short time interval in the same group might be related to the same conversation. We emphasize that by no means does the elapsed time category imply that a message that appeared more than an hour after the last message was sent in a given group is not related to former messages, except that with no other supporting data (i.e., the content itself) it is impossible to make a concrete connection to prior messages. Hence, the time elapsed is the only possible, though not a perfect, indication for relevance. We also discretized messages according to their file attachments and created Boolean categories of messages with and without files.

The second type of logged information concerned WhatsApp conversation groups. This dataset contained a total of 10,730 such groups from the 111 users. Note that groups with two participants are similar to a typical SMS conversation, and thus through logging this data we could test the degree to which WhatsApp has replaced traditional SMS messaging. However, groups might also be formed around a general topic, such as a discussion about work, leisure, or family issues with many more than two participants. We logged information about the group size of all of the messages and categorized this information into the percentage of messages in trivially small groups of two people, groups of three to four participants, and those with five or more participants. We also collected group statistics that subsume those within the message analysis, but we refer to the percentage of messages within a group having a certain attribute (e.g., the percentage of messages sent at a certain time, of a certain length, containing a file, etc.).

4.Descriptive statistics

The general methodology assumption behind this paper is that the analysis must be datadriven. Therefore we use the data to support any assumptions about the nature of the data. In contrast, previous studies typically assume some type of behavior and then construct questionnaires to prove or disprove that assumption (Church and de Oliveira 2013; Pielot et al. 2014; O'Hara et al. 2014; Mudliar and Rangaswamy 2015). For the data- driven approach to be successful, significant differences must be evident across different demographic groups within the data. To confirm this assumption, we checked that such differences did in fact exist and were statistically significant.

Specifically, we analyzed the basic distribution of messages, focusing on the statistical distributions across different genders, ages, and types of use. We found that over 70% (71.5%) of WhatsApp groups had only two participants (7,671 out of the 10,730), confirming previous assertions that WhatsApp is replacing SMS messaging (Church and de Oliveira 2013). On the flipside, over 50% of all messages were not in groups of two (3,713,052 out of 6,449,631), indicating that, typically, larger groups were fruitful grounds for larger discussions - something that SMS typically does not support. To better understand this point, please note these differences using the graphical distributions of the number of groups of each size in Figure 1 and the distribution of all messages in those same groups in Figure 2. We note that the number of two-person groups is overwhelmingly large (71.5%), but that the number of messages in these groups is significantly smaller (42.43%).

Notably, we found that while the number of groups with over 50 members is less than 1%, these groups have a disproportionately large number of messages (8.37%). We believe that the reason for this is clear: Larger groups tend to have larger numbers of messages in each group. Thus, we find that a large percentage of WhatsApp activity is in fact taking the place of traditional SMS messages between two people. However, group messaging among large numbers of users, another key use of WhatsApp which SMS is less successful in supporting, also constitutes a large percentage of the WhatsApp messages we collected.

We then studied the statistical distribution of the messages' attributes, starting with the average response time (the time that elapsed between any two messages in a conversation), which can be found in Figure 3. Please note that the average response time is quite short. Over one half (57.82%) of all messages are responses that were composed within one minute. This finding again confirms previous claims that WhatsApp has become a replacement for traditional SMS messaging because most participants answer their messages quite quickly - something that is expected with SMS messaging (Church and de Oliveira 2013). We also analyzed the message types sent. We found that most of the messages (approximately 99%) were exclusively text messages, while only 1% included file attachments or links. Last, we studied the distribution of the messages throughout the day (which is visually represented in Figure 4). As expected, very few messages were sent overnight, with under 5% (4.36%) being sent between midnight and 4:00 a.m. and only 2.37% being sent between 4:00 a.m. and 8:00 a.m. Note that fewer messages were sent between 8:00 a.m. and noon (18.04%) compared to approximately 25% of all messages being sent in each of the other four-hour intervals. In fact, we note no significant difference in the number of messages being sent in these three intervals (p- score > 0.1), while a significantly smaller number of messages were sent between 8:00 a.m. and 12 p.m. (p-score << 0.01).

Table 1 contains several additional gender-related insights. First, we found that women on average sent and received more messages than men. Women sent and received over 155 messages a day, whereas men sent and received approximately 134 messages (row 4), a difference of approximately 15%. Of these messages, women sent on average approximately 46 messages a day and received 109 messages, whereas men sent an average of slightly less than 30 messages a day and received about 104 messages (rows 2-3). Thus, men evidently sent fewer messages on average than women, something which was also evident in the differences in the ratios between sent and received messages (row 5). Second, while on average both genders participated in a similar number of conversations overall (63-65 groups), the distribution of the various group sizes between the genders was different. Women were more active in smaller conversation groups (60.13% versus 55.31% in groups with two participants), whereas men were more active in larger groups (11.05% for men versus 6.52% for women in groups of 11-20, 8.55% versus 5.91% in groups of 21-50 and 2.13% versus 0.75% in groups bigger than 50) (rows 6-12).

We note which of these differences proved statistically significant by using a twotailed t-test. Cells that are annotated with one star (i.e., AvgSent/Day) recorded a p-score of under 0.1 (0.06), while those cells with two stars were below the 0.05 significance level (0.02, 0.01, and 0.04 for the RatioSend/Rcv, groups sized 11-20 and 51+ respectively). Thus, we note that significant differences do exist between the genders' behavior.

We also found that there were significantly different WhatsApp usage patterns between different genders and age groups. Table 2 provides details to support this claim. We present the general statistics of two different demographic groups: (1) men and women and (2) WhatsApp users younger than 25 (the median age) and aged 25 or older. We selected these age distributions based on a previous large-scale statistical analysis of WhatsApp user ages in the general population (http://www.statista.com/statistics/290447/agedistribution-of-us-whatsapp-users).7 Note the differences between the average number of total messages per day (AvgMsgDay), groups per user (AvgGroup/Usr) and differences in the users' responses to the questionnaire items in which they self-rated their sociality (SocialLevel), overall usage (UsageLevel), differences in the Boolean values (averaged based on values of 0 and 1), and usage in communicating with friends (UsageFriend), family (UsageFamily), and work (UsageWork). In fact, we tested all pairs of numbers for statistical significance (two-tailed t-test) and found that all differences were significant (p-score << 0.05) except where noted with a "#" at the end of each pair, as is the case of the UsageWork numbers in the pair of people 25 or older and younger than 25. Additionally, we found significant differences in the usage patterns across group usage with people who were members of these different demographics. Note the differences in the average number of minutes a user took to respond to a message (AvgResponse), the percentage of their messages that were five words or fewer (Msgs5orLessWrd), the percentage of their messages that were quick responses within five minutes (%RespUnder5), the average message length (AvgTextLength), and the distribution of messages across different times (midnight to 4:00 a.m., 8:00 a.m. to 12:00 p.m., and 8:00 p.m. to midnight). We also found that usage styles were different in regards to the percentage of files found in users' groups of different genders and ages (UseFile) and the percentage of groups of which they were members that had five or more total users (isGrp5+).

We find some of the differences in Table 2 intuitive and others surprising. We are not surprised to find that younger people are more likely than older ones to send messages late at night, and thus relatively older people send a higher percentage of their messages during the day. One could find support for gender differences in people's self-rating of how much they use WhatsApp to communicate with family versus work, similar to previously observed differences in gender expressions (Kring and Gordon 1998). However, we could not find a clear explanation as to why men seem to send more files in their groups than women or why older people participate in larger groups more often than younger people. These differences might point to new directions that could be confirmed with further research and questionnaires. For example, a possible hypothesis for the differences in group sizes across different ages is that younger people have more thoroughly adopted WhatsApp as a replacement for SMS messaging and consequently a larger percentage of their communication can be found in these smaller groups.

5.Predictive models and hypotheses

As we demonstrated in the previous section, significant differences do in fact exist between different types of WhatsApp users and groups. However, even statistically significant differences do not necessarily allow us to predict usage patterns. For example, the previous section demonstrated that men typically send shorter messages and women send and receive more messages per day. However, these differences do not necessarily allow us to make a prediction about a specific user - something that data mining algorithms do in fact allow, as we now present. To illustrate the potential of using the collected data for prediction purposes, we created several predictive models for the user and group datasets, which we describe in this section.

User models were based on the 111 users in this dataset and were built to identify whether the author of a given set of WhatsApp posts is of a given gender or age. Our first hypothesis is that differences between WhatsApp users can be predicted by exclusively using general statistics about usage, even without specific user content. In accordance with the results reported in the previous section, we posit that such differences will probably use attributes such as message length and response time because such attributes might be affected by known gender differences (Kring and Gordon 1998). For example, one might find that women write more to better express their ideas or emotions, while men write more curtly. Similarly, one might find that differences in response time or average conversation length reflect emotional difference - e.g., women may prefer discussions in small groups while men prefer less personal, larger discussions. In a similar vein, one might find differences between ages, even within one gender. Such differences may be somewhat obvious, such as the time at which a message is sent - e.g., people of certain ages might be more or less likely to work and thus be less likely to send messages at certain times. But less obvious differences might exist too, such as differences in message length.

Our second hypothesis, based on the differences reported in accordance with the various statistics described in the former section, is that different types of group usage can be predicted based on general group attributes, again even without considering the messages' content. Specifically, we develop models that predict which groups will have a certain type of content such as file attachments or shorter messages. We also develop group models that predict which groups will have certain user activity, such as a larger quantity or more frequent messages, and quicker response times. In theory, other usage questions could have been studied, such as if a message contained certain text - e.g., inappropriate or flagged for a certain type of content. However, because we have no access to message content, these issues cannot be evaluated. Similarly, it may be possible that certain messages are inherently different and thus likely to be more popular or important. Along these lines, models might be created to predict which messages are apt to have certain characteristics, such as being forwarded - something that was previously studied within the Twitter network (Naveed et al. 2011). However, once again that study focused on the message content, which is often infeasible to rely on in real-life settings, due to either privacy constraints or availability.

The advantage of using data mining algorithms to test these hypotheses is the objectivity of the results. On a technical level, we built models from decision trees, as implemented in the C4.5 algorithm (Quinlan 1996) to create classifiers between two choices (Boolean). The C4.5 algorithm was chosen because of two main advantages. First, C4.5 identifies which attributes are most important for accurate prediction by using the Info-Gain measure to rank the predictive ability of all attributes. This allows us to objectively identify which factors are most important for accurate prediction. Second, the if-then rules output by these algorithms allow us to observe and analyze the exact range of values within the selected attributes that form the prediction model. Furthermore, we consider many tasks - such as if a user is male or female, or above or below a certain age - which are inherently Boolean decisions and are thus well suited for C4.5. To handle continuous attributes, we transformed the target variables into two categories through binning according to preset cutoff thresholds. For example, in creating the quick-response-time model, we chose a response threshold of one minute. We then created a Boolean classifier and assumed that anyone who answered within 1 minute answered quickly and those who answered after one minute, even if they answered only seconds after one minute, did not. More specifics of the models and their findings are in the next section.

6.Data analytic results

In general, we built two types of models using the popular open source Weka data mining package (Witten and Frank 2005): decision trees and probabilistic models based on Bayesian networks. We did consider other models, but the Bayesian models often did better than the other alternative algorithm, so we present results from this algorithm for comparison. Within the user models, standard ten-fold cross-validation was used to assess all models because the validation set was always a different set of users than those used in the training data. While we also considered using standard cross-validation in order to assess the group prediction models, we rejected this approach because sometimes we noted that both the training and testing datasets contained groups from the same user. Instead, for each group model we generated ten randomized splits which ensured that a user's groups were within only the training or the testing dataset. While the resultant stratified training-testing splits were not always of the same size, they did guarantee that we did not overfit by having the same user in both the training and testing data.

We also created logistic regression models using the pROC package in R (Robin et al. 2011). While these models were not validated using cross-validation, we did generate confidence intervals at the 95% level. These models provide additional confirmation and comparison of our results to a linear model, including the statistical analysis provided by the confidence intervals.

Overall, we were successful in predicting a user's gender and approximate age based on users' general data, as can be seen in Table 3. The first column in the table presents how many records were in each dataset. The second column represents a baseline accuracy which, is constructed by classifying all records as belonging to the larger category (referred to as ZeroR in Weka). For example, the user dataset contains 59 women and 52 men. Assuming all users are female would have an accuracy of 53.15% (see column 2, row 1) and 63 of the 111 users were younger than 25, thus leading to a baseline value of 56.76% (see column 2, row 2). Minimally, a successful model should at least be more accurate than this value. The next two columns present the accuracy and area under the curve (AUC) of the decision tree model, with the following two columns presenting the accuracy and AUC of the corresponding Bayesian model. The last two rows present the AUC and confidence intervals at the 95% level for the logistic regression model in the pRoc package in R. The first row presents the results for predicting gender based on the data and the second row presents the results for predicting age - i.e., 25 or older versus under 25 years old. This cutoff was chosen because it represented roughly a 50/50 split within the data. Note that both models were successful in both tasks: The predictions' accuracies were much greater than the baseline values.

An advantage in building decision trees is noting the logical rules and the attributes in the learned models. We studied this output from the decision tree models. We noted that the decision tree for predicting gender focuses on the usage level, the response time and the number of large groups a user had. Specifically, we found that men overall selfreported lower usage levels, on average took longer to respond, and have large groups. A slightly simplified version of the gender decision tree is presented in Figure 5. Note that this rule is relatively simple: If the user self-rated a usage level of 3 or less, they were male, otherwise they were female. Despite the simplicity of this decision tree, it still yielded an accuracy of 63.56%. Nonetheless, more complex models could be built with both decision trees and Bayesian networks as reported in Table 3.

Using a similar methodology, we were able to differentiate between users below the age of 25 and those above it. Here, we noted that younger people had more messages a day (AvgMsgDay) and were likely to be in groups with more than five people (isGrp5+). The decision tree for predicting age, found in Figure 6, shows the exact rules behind this classifier. The model predicted that if a person had more than 1.33% of all groups belonging to a group of 50 or more members then they were under 25, but if they had fewer than this number of large groups and received on average of fewer than 117 messages per day they were 25 or older. Otherwise, a third rule was needed to differentiate between younger people, with more than 8.62% of their groups constituting five or more people, and older people, with fewer groups of this size. We again note that while these rules are consistent with the general trends seen in Figure 2, the decision tree provides a predictive model with exact thresholds that predict differences between the groups.

We also created models for the gender and age prediction tasks using the group dataset, the results of which are in the first two rows of Table 4. Overall we have much more group data (10,730 records) than average user data (111 users), so it is not surprising that these models often performed better than the user models, particularly when noting the differences in AUC. Again, the decision trees provide insight as to which attributes are most helpful. In the first decision tree we found that once again men were characterized by lower usage levels, used WhatsApp less for family communication, and had shorter messages. Within the age classification task, we found that younger people sent files less frequently than older people and were less likely to use WhatsApp for family and work communication. While the attribute AvgMsgSent played prominently in the classification task from the user dataset, this attribute was absent from the group dataset because average statistics for a user are not evident from a group's profile. Nonetheless, the group statistics proved to be even more helpful in building age and gender models.8

We also built models that predicted group usage characteristics, the results of which are also found in Table 4. Specifically, we built models to predict which groups will contain file attachments in at least 1% of all messages (row 3), which participants will have more than 25% of their messages sent between 5:00 a.m. and 9:00 a.m. (row 4), which messages are characteristic of groups with five or more users (row 5), which participants will average at least a total of five messages sent or received per day (row 6),which messages will on average contain short texts with five or fewer words in at least 75% of all messages (row 7), and which participants will receive responses to at least 25% of all messages within one minute (row 8). In general, these models were much more successful than the baseline values, with AUC values often above 0.8. However, some exceptions do exist. Note that predicting age based on group activity was not successful within the decision tree model with an AUC value of below 0.5 (0.36). Nonetheless, even here the Bayesian model was more successful, with an accuracy nearly 30% greater than the baseline and an AUC above 0.5. The logistic regression models based on all data are meant to provide another comparison. Notably, the confidence intervals point towards the models here being general (i.e., even the lower bound of the AUC confidence intervals are always above 0.7).

The thresholds used in this task were meant to be representative of groups that are more active in the morning and have shorter messages and shorter response times. We did in fact check other tasks, such as classifying the prediction of which messages were sent later at night and in larger groups. We also created models considering different thresholds for the group sizes and response times, as well as for other tasks. Overall, we found that the data similarly supported prediction beyond the specific thresholds reported in this paper. Important future work includes considering additional tasks and thresholds beyond those considered in this paper.

The decision tree models also facilitated the ability to understand which attributes were most influential in predicting the group's behavior. For example, a simplified version of the decision tree to predict which groups had one quarter of the messages answered within a minute is found in Figure 7. Specifically, groups with a larger percentage of shorter messages (1-2 words and 3-5) typically indicate quick answers. Similarly, we were able to use decision trees to understand the models for other group behaviors. We found that groups with more files were typically composed of younger participants (28 or younger) with advanced schooling (a master's degree or more). Additionally, groups with users who didn't rate themselves with high usage levels but had high educational levels (above 16 years) and were above 30 still typically sent more file attachments. As one might expect, we also found that full-time students were less likely to be active in the morning compared to those who had jobs. As Figure 2 demonstrates, we also found that groups with five or more participants have more messages and thus typically have messages sent with a higher frequency. We found that larger groups typically contained shorter messages. We also found that younger people typically send shorter messages, and while older people typically send longer messages, they do so less frequently.

7.Conclusions and future work

This work represents the first exhaustive analysis of WhatsApp messages. We collected over 6 million encrypted messages from over 100 students between the ages of 18 to 34, and differentiated between different types of user and group use of the network. A key characteristic of this study is that storing encrypted messages prevented us from analyzing any content within the messages. This was done intentionally to safeguard participants' privacy. Despite this limitation, we found that many message and group characteristics significantly differed across users of different demographics, such as gender and age, and present these results through performing extensive statistical analysis. Additionally, we believe that one key novelty of this work is that we use data analytics to predict the users' gender, age, and group activity. Our work is data driven, so we based our findings on the algorithms' output. We did not attempt to verify any specific thesis, as had been done previously. This is one key advantage of using data analytics, and this difference is especially clear from the decision tree results presented in this paper.

Overall, our results provide several new insights into WhatsApp usage. We find that the younger users in this dataset used this network more frequently. We also find that more years of education and age are positive factors in predicting how frequently people send file attachments. Overall, women use this network more often than men, and they reported that they use it more often to both generally communicate and to communicate with family. Men, on the other hand, are generally members of larger communication groups and send shorter messages. Additionally, larger groups are defined not only by their large number of users, or even the large numbers of messages that are frequently sent, but also are typically defined as having shorter messages than those in private oneto-one communications. Decision tree models were not only helpful in identifying these attributes but were useful in providing the thresholds within the if-then rules for the models that predicted these results. Because our results are built through analyzing users' general message data, but without message content, we believe that the methodology used in our analysis may be of general interest to other groups, such as demographers and government bodies to facilitate data analysis without infringing on users' privacy.

In building upon this work, we believe that two types of studies will probably lead to fruitful results. First, we believe that additional studies should be undertaken to improve upon and extend the study we present. While this study analyzed over 6 million messages, it is still limited in containing only 111 users and focusing exclusively on people between the ages of 18 and 34. Furthermore, we believe it will be helpful to study how different demographic groups use WhatsApp. We believe that even more accurate models can be built through studying data from more users, with a wider range of ages and different ethnic backgrounds. Similarly, we did not study all group tasks, and other tasks - such as which messages will be forwarded - remain unexplored. We also did not consider all possible thresholds within the tasks studied, such as the percentage of messages answered within five minutes or 30 seconds, in contrast to the threshold of one minute presented in this paper. In a related matter, while we intentionally built models without analyzing user content in order to safeguard privacy, even more accurate models might be built in the future if user consent could be obtained for this information. We are hopeful that researchers will build upon this study and perform additional analyses both within this dataset and other, possibly larger, WhatsApp datasets.

We believe a second type of direction should focus on applying the lessons learned from this paper's models. It may be wise to customize user interfaces for certain types of users and tasks based on the attributes found to be important in this paper. For example, users who are more educated or older might prefer a different WhatsApp interface compared to less-educated or younger users, given that their usage patterns differ significantly. Similarly, since larger groups are characterized by shorter messages, it may be that the interface for these types of interactions should be customized with this information in mind as well. We hope that these and other issues will be explored in greater detail in future work.

Footnote

3http://www.wired.com/2016/02/one-billion-people-now-use-whatsapp.

4While we note that a computer interface for WhatsApp exists, it is exclusively an interface for people's smartphones and offers no additional functionality.

5 Both the ADB software and the data collected are available from the authors.

6 The software we used collected all the data on the phone, hence the time period over which data was collected varied according to when users started using WhatsApp and their habit of deleting old messages (if at all).

7 We leave analysis of different age groups for future studies.

8 It is interesting to note that the baseline for the age group here is only 33.19% or worse than random selection. This is because the random stratification typically chooses a different majority condition in the training and testing cases here (e.g., a majority of records with an age under 25 in the training but a majority of cases over 25 in the testing dataset. This may explain why both the baseline and decision tree models here are relatively poor.)

References

References

Argamon, S., Koppel, M., Pennebaker, J.W., and Schier, J. (2009). Automatically profiling the author of an anonymous text. Communications of the ACM 52(2): 119-123. doi:10.1145/1461928.1461959.

Bakshy, E., Rosenn, I., Marlow, C., and Adamic, L. (2012). The role of social networks in information diffusion. In: Mille, A., Gandon, F.L., Misselis, J., Rabinovich, M., and Staab, S. (eds.). Proceedings of the 21st International Conference on World Wide Web. New York: ACM: 519-528. doi:10.1145/2187836.2187907.

Bouhnik, D. and Deshen, M. (2014). WhatsApp goes to school: Mobile instant messaging between teachers and students. Journal of Information Technology Education: Research 13: 217-231. doi:10.28945/2051.

Church, K. and de Oliveira, R. (2013). What's up with WhatsApp? Comparing mobile instant messaging behaviors with traditional SMS. In: Rohs, M., Schmidt, A., Ashbrook, D., and Rukzio, E. (eds.). Proceedings of the 15th International Conference on Human-Computer Interaction with Mobile Devices and Services. New York: ACM: 352-361.

Fan, W., Wallace, L., Rich, S., and Zhang, Z. (2006). Tapping the power of text mining. Communications of the ACM 49(9): 76-82. doi:10.1145/1151030.1151032.

Fiadino, P., Schiavone, M., and Casas, P. (2014). Vivisecting WhatsApp through largescale measurements in mobile networks. Paper presented at SIGCOMM, Chicago, USA, August 17-22, 2014. doi:10.1145/2619239.2631461.

Gulacti, U., Lok, U., Hatipoglu, S., and Polat, H. (2016). An analysis of WhatsApp usage for communication between consulting and emergency physicians. Journal of Medical Systems 40(130): 1-7. doi:10.1007/s10916-016-0483-8.

Jain, J., Eddy Luaran, J., and Rahman, N.b.A. (2016). Learning beyond the walls: The role of WhatsApp groups. In: Luaran, J.E., Sardi, J., Aziz, A., and Alias, N.A. (eds.). Envisioning the future of online learning: Selected papers from the International Conference on e-Learning 2015. Singapore: Springer: 447-457. doi:10.1007/978-981-100954-9-40.

Johnston, M., King, D., Arora, S., Behar, N., Athanasiou, T., Sevdalis, N., and Darzi, A. (2015). Smartphones let surgeons know WhatsApp: An analysis of communication in emergency surgical teams. American Journal of Surgery 209(1): 45-51. doi:10.1016/j.amjsurg.2014.08.030.

Kring, A.M. and Gordon, A.H. (1998). Sex differences in emotion: Expression, experience, and physiology. Journal ofPersonality and Social Psychology 74(3): 686-703. doi:10.1037/0022-3514.74.3.686.

Lin, Y.H., Lin, Y.C., Lee, Y.H., Lin, P.H., Lin, S.H., Chang, L.R., Tseng, H.W., Yen, L.Y., Yang, C.C.H., and Kuo, T.B.J. (2015). Time distortion associated with smartphone addiction: Identifying smartphone addiction via a mobile application (App). Journal of Psychiatric Research 65: 139-145. doi:10.1016/j.jpsychires.2015.04.003.

Montag, C., Błaszkiewicz, K., Sariyska, R., Lachmann, B., Andone, I., Trendafilov, B., Eibes, M., and Markowetz, A. (2015). Smartphone usage in the 21st century: Who is active on WhatsApp? BMC Research Notes 8(1): 1-6. doi:10.1186/s13104-015-1280z.

Mudliar, P. and Rangaswamy, N. (2015). Offline strangers, online friends: Bridging classroom gender segregation with WhatsApp. In: Begole, B., Kim, J., Inkpen, K., and Woo, W. (eds.). Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems. New York: ACM: 3799-3808. doi:10.1145/2702123.2702533.

Naveed, N., Gottron, T., Kunegis, J., and Alhadi, A.C. (2011). Bad news travel fast: A content-based analysis of interestingness on Twitter. In: De Roure, D. and Poole, M.S. (eds.). Proceedings of the 3rd International Conference on Web Science. New York: ACM: 8:1-8:7. doi:10.1145/2527031.2527052.

O'Hara, K.P., Massimi, M., Harper, R., Rubens, S., and Morris, J. (2014). Everyday dwelling with WhatsApp. In: Fussell, S.R., Lutters, W.G., Morris, M.R., and Reddy, M. (eds.). Proceedings of the 17th ACM Conference on Computer Supported Cooperative Work and Social Computing. New York: ACM: 1131-1143. doi:10.1145/2531602.2531679.

Pielot, M., de Oliveira, R., Kwak, H., and Oliver, N. (2014). Didn't you see my message? predicting attentiveness to mobile instant messages. In: Jones, M., Palanque, P.A., Schmidt, A., and Grossman, T. (eds.). Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. New York: ACM: 3319-3328. doi:10.1145/2556288.2556973.

Quinlan, J.R. (1996). Improved use of continuous attributes in C4.5. Journal of Artificial Intelligence Research 4: 77-90. doi:10.1613/jair.279.

Robin, X., Turck, N., Hainard, A., Tiberti, N., Lisacek, F., Sanchez, J.C., and Muller, M. (2011). pROC: An open-source package for R and S+ to analyze and compare ROC curves. BMC Bioinformatics 12: 77. doi:10.1186/1471-2105-12-77.

Thelwall, M., Wilkinson, D., and Uppal, S. (2010). Data mining emotion in social network communication: Gender differences in MySpace. Journal of the Association for Information Science and Technology 61(1): 190-199. doi:10.1002/asi.21180.

Van Wel, L. and Royakkers, L. (2004). Ethical issues in web data mining. Ethics and Information Technology 6(2): 129-140. doi:10.1023/B:ETIN.0000047476.05912.3d.

Wagner, C., Garcia, D., Jadidi, M., and Strohmaier, M. (2015). It's a man's Wikipedia? Assessing gender inequality in an online encyclopedia. Paper presented at the 9th International AAAI Conference on Web and Social Media, Oxford, UK, May 26-29, 2015.

Wang, Y.C., Burke, M., and Kraut, R.E. (2013). Gender, topic, and audience response: An analysis of user-generated content on Facebook. In: Mackay, W.E., Brewster, S.A., and Bødker, S. (eds.). Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. New York: ACM: 31-34. doi:10.1145/2470654.2470659.

Witten, I.H. and Frank, E. (2005). Data mining: Practical machine larning tools and techniques. Burlington: Morgan Kaufmann.

Xiang, R., Neville, J., and Rogati, M. (2010). Modeling relationship strength in online social networks. In: Rappa, M., Jones, P., Freire, J., and Chakrabarti, S. (eds.). Proceedings of the 19th International Conference on World Wide Web. New York: ACM: 981-990. doi:10.1145/1772690.1772790.