Twitter is a popular, free-to-use microblogging social media platform on which users can instantly broadcast short messages. These short messages are called tweets. Tweets are limited to 280 characters (this changed from 140 characters in November 2017) and are meant to be broadcasted to the world. Twitter users, however, can also send out directed, conversational messages by using the @ symbol followed by a Twitter user’s ID. Twitter does not have topically focused groups, but Twitter users can use the # symbol to label and categorize their tweets with searchable keywords. We employed the Python library Tweepy [49] and used hashtags from a previous study [50], including #weightloss, #diet, #fitness, and #health, to collect tweets related to physical activity and weight loss. We then extracted motivation-related tweets to answer RQ1 and reciprocal conversational interactions to answer RQ2. In this study, we only examined tweets in English.

We first used a lexicon-based approach, which is a method to identify related text passages based on a set of words or phrases, to extract and examine motivation-related tweets from the overall physical activity tweets. A number of different motivation-related terms were considered, but for this study, we only used a single term, motivation, for the following three reasons. First, Twitter users frequently use acronyms and abbreviations due to the 280-character length limitation. However, in an online environment, these terms have been shown to change over time [51]. The same acronyms and abbreviations can also be used differently in different communities, which have specific cultures [52]. Second, we placed a higher priority on precision than recall in order to accurately observe user behavior and reduce noise. Third, many popular motivation-related hashtags contain the term motivation (eg, #fitnessmotivation and #mondaymotivation). Thus, we filtered for tweets that contained the keyword motivation. Specifically, we first preprocessed the entire data set to convert the text to lowercase. Then, to extract tweets containing our key terms from the entire data set, we employed a lexicon-based approach and extracted tweets and data from their “user” and “entities” fields, including user names, screen names, locations, timestamps, and user mentions. We extracted and included any partial matches in this process to cover a wide variation in terms, such as #mondaymotivation.

To understand temporal trends in motivation-related messages, we first analyzed the data by week and by time of day. In this analysis, we included data from between September 1, 2018, and November 2, 2018, because the tweets in our original data had an uneven number of days of the week and were collected at different starting and ending time points. The hours of the day were grouped into eight periods: (1) early morning (4:00 AM to 6:59 AM), (2) morning (7:00 AM to 9:59 AM), (3) late morning (10:00 AM to 12:59 PM), (4) afternoon (1:00 PM to 3:59 PM), (5) evening (4:00 PM to 6:59 PM), (6) late evening (7:00 PM to 9:59 PM), (7) night (10:00 PM to 12:59 AM), and (8) late night (1:00 AM to 3:59 AM).

To characterize and understand these messages, we first qualitatively examined retweets—reposted or forwarded tweets—related to motivation. Retweets are important in Twitter, because they have a higher chance of reaching a large number of users and can represent messages that embody the collective ideation of the platform [53]. Thus, we qualitatively analyzed the general themes of retweeted tweets following the thematic analysis process [54,55]. A prespecified threshold of at least 50 retweets was applied for a tweet to be analyzed, due to the size of the data set. All of the quotes in this paper have been deidentified and slightly modified to protect the privacy of users following the guidelines suggested by a previous paper [56].

To quantitatively expand the retweet analysis, we created a topic model using latent Dirichlet allocation (LDA) [57] to identify main topics from the larger data set. Topic modeling is an effective method to discover topics in a large data set and can identify a set of topics from a given set of documents based on document-level word co-occurrences. In our study, each tweet was considered a single document. First, we preprocessed the data set by removing URLs and tweets with fewer than 5 words. Similar exclusion criteria were applied to reduce noise in a previous study that used social media data [58]. Then, nouns were extracted using the Python natural language toolkit library [59] and used in the topic modeling process. LDA requires a predetermined number of topics, which can be a limitation. After experimenting with varying numbers of topics, we chose 10 topics related to motivation and physical activity on Twitter. An LDA topic model was generated by the Python library genism [60], followed by manual examination and labeling of the identified topics. To visually review these topics and occurrences of associated terms, we then visualized the topic model (ie, the main topics and their top 50 associated words) as a word cloud using the Python library wordcloud (Multimedia Appendix 1) [61]. Word cloud visualization is simple and user-friendly, yet scalable and preferred by users [62,63].

We generated 2 social network visualizations to verify whether social media enables organically formed peer support pairs or groups for motivating physical activity and to examine their social network structure. The first Twitter social network visualization represents a social structure of reciprocal, conversational interactions that are captured when users direct their tweets using the @ symbol to other Twitter IDs. These interactions are dyadic ties among Twitter users (determined by Twitter ID), who are the social actors. We first extracted all tweets directed toward other users by checking if the @ symbol was placed in conjunction with a Twitter ID in a tweet. We then use Gephi [64], a popular network visualization tool, to visualize the social network and apply the OpenOrd [65] layout to gain an overview of the network structure. The edge weight is determined by the number of interactions among Twitter users.

To examine Twitter’s real-world potential for connecting physical activity partners [38], we generated a second social network visualization of users who disclosed their location information. The purpose of this social network visualization was to visually examine the prevalence of social interactions that are generated from the same physical location and to assess the relationship between users’ current practices and the platform’s potential utility for connecting physical activity partners. Twitter users can opt to share their location information in their user profile. The Twitter API also provides a mechanism to track the geocode of a tweet; however, a previous study suggests that only 1.5% of tweets are geotagged [66]. Thus, we extracted users’ location information from their profile for the second social network visualization and also recorded the number of geotagged tweets and users who opted to disclose this information. We removed social actors without location information, and then estimated peer-supporting physical activity partners (ie, socially and physically connected users). We denoted this information using a coloring schema in the second social network visualization. We used the Fruchterman Reingold graph layout algorithm [67] for the second visualization, which had a network structure with considerably fewer social actors and dyadic ties.

The study was determined to not have human subjects by the University of North Carolina-Charlotte's Institutional Review Board (Ethics Committee).

We collected over 1.5 million physical activity–related tweets made between August 30, 2018, and November 6, 2018. Table 1 summarizes the overall data set and the data sets for each research question.

The overall social network of directed tweets consisted of 14,028 dyadic ties among 10,889 social actors (determined by Twitter ID). It is important to note that some tweets contained multiple indications of where they directed their tweets, included self-mentioning, or were directed to Twitter IDs that were not found in our data set. This resulted in a greater number of social actors than the total number of Twitter IDs but a lesser number of dyadic ties than the total number of tweets shown in Table 1 and Table 3. Table 3 summarizes the directed tweets and the creators of those tweets (determined by Twitter ID) in our data set. About 68% (4459/6558) of the Twitter users disclosed their location in their profile; however, users rarely disclosed their exact location by providing geocode information. Similar to a previous study [66], our data set consisted of about 1% (94/6558) geotagged tweets and Twitter users, resulting in less than 100 users who opted to disclose their exact location.

Figure 1 shows an overview of the Twitter social network in our data set. Twitter users formed hundreds of small groups that sent motivation-related tweets to each other, while a small number of larger groups with a higher number of dyadic ties and social actors were formed. Figure 2 shows a social network of location-based Twitter user interactions. After removing social actors without location information, self-mentioning tweets, and interactions with only one-sided location information, the social network structure consisted of 2878 dyadic ties among 1112 social actors, a smaller number of social actors and dyadic ties than shown in Table 3. Of these 2878 social interactions, 1.88% (54/2878) were made from the same physical location.

A relatively short timeframe and a lack of seasonal coverage are limitations of our study. Though we did not examine seasonal trends, it is possible that posting frequency could change with the season. Similarly, previous work has shown that social media usage can vary temporally and spatiotemporally [83]. The time period in our study may also have affected trends in posting. Our study time period included a change of seasons, from late summer to fall, which may have been associated with changes in activity and exercise. Similarly, we manually examined nearly 11.77% (1562/13,268) of the total retweets, but the number of retweeted original tweets in our qualitative analysis was relatively small (18/5915). Although we used a prespecified threshold to select original tweets to be qualitatively examined, investigating a larger number of original tweets could have led to more conclusive findings. Moreover, future studies should examine long-term engagement with content among group members to understand the formation, evolution, and collapse of peer networks for the purpose of encouraging physical activities.

Twitter has a diverse user group compared to many other social media platforms, although it is still used more by young adults [84]. This could have created a bias toward younger Twitter users, allowing us to identify particular topics of interest only to this group. We do not know how our findings would apply to other demographics or social media sites. Similarly, our analysis focused on English tweets, and our findings may not be generalizable to non–English speaking countries. Our findings, however, have significant implications for young adults, the main user group of social media, and the health of future society, as they develop lasting health-related behaviors.

Another selection bias was present in the location-based Twitter user interaction analysis (Figure 2). Although the majority of users opted to disclose their location information in their profile (4459/6558, 67.99%), this analysis was conducted with users who self-selected to disclose this information. In the same analysis, an additional limitation was discovered in our manual validation check. We believe that the different location figure (2824/2878, 98.12%) was inflated due to the granularity of information provided by users. For example, in our data set, we found a few users who had identified their location as “Earth” or “America.” Moreover, some users disclosed their information at the state level (eg, North Carolina) and some users disclosed this information at the city level (eg, Charlotte). Although it is possible that they such users were in fact at the same physical location, we were not able to determine this without additional information. In our study, they were thus treated as being in different locations.

In this study, we used hashtags that were identified in a previous study that collected tweets related to physical activity for the purpose of weight loss. However, we did not incorporate postprocessing to include a wider variation in physical activities. The collected tweets were then processed to extract motivation-related tweets for RQ1. Initial key terms are important in collecting a relevant data set, and different sets of terms could have resulted in different findings, especially related to the prevalence of these tweets. Similarly, for RQ1, we used a single keyword, motivation, to identify motivation-related tweets. While this approach can identify related tweets with high precision, we should in future extend these study results to retrieve a greater number of relevant messages by developing a classifier. Moreover, major differences, such as communication style, user demographic, and the platform’s intended purpose, exist among social media platforms. To better characterize the landscape of peer support and motivating messages for physical activity in social media, replicating a similar study on a different popular social media platform, such as Facebook or Reddit, is warranted.

Peer support through social media could be a potential means of encouraging positive behavior changes and sustaining that changed behavior. This study examines whether individuals organically use social media to encourage and sustain physical activity. The organic exchange of motivational messages and peer support are found among Twitter users, although this support is more commonly used by individuals from different physical locations. Our findings suggest that geographical location is less important in providing peer support, as long as the support provides motivation, despite the users having few factors in common affecting their physical activity (eg, the weather). Based on existing interaction patterns, we consider that successful motivation-providing peer support groups or interventions could be delivered through social media. However, further research on the effects in a real-world context, as well as additional design and usability features improving user engagement, are warranted to develop a successful intervention to counteract the current obesity pandemic. This is especially important for young adults, the main user group of social media, as they develop lasting health-related behaviors.