More than Meets the Tie: Examining the Role of
Interpersonal Relationships in Social Networks

Linguistic style and content reflect how an individual perceives another (Bell 1984). For instance, usage of pronouns reveal levels of self-disclosure (Choudhury and De 2014; Wang, Burke, and Kraut 2016), and swearing terms indicate a closer social distance between the speakers (Feldman et al. 2017). Comparing the use of these words by relationship category can reveal how open each relationship types are in Twitter conversations. Using lexicons from LIWC (Pennebaker et al. 2015), we calculate the probability of a directed mention containing one of a specific set of words: \nth1 person singular and plural pronouns, \nth2 person singular pronouns, and swearing terms. Assuming that there exist topics central to a single type of relationship such as work-related topics, we also include the LIWC categories for work- and family-related words. The results are displayed in Figure 1.

Communication patterns match prior expectations, as illustrated in three trends. First, conversations in organizational relationships focus on collective identity, as shown in the highest probability of \nth1-personal plurals, but lowest in \nth1-person singular, as individuals in these relationships associate each other in the context of a larger collective entity (Klein and D’Aunno 1986). Second, parasocial relationships use lesser \nth1-person plural pronouns but more \nth1- and \nth2-person singular pronouns, making their behavior similar to that of romantic relationships. This result is consistent with previous findings on behaviors in parasocial relationships that resemble love and affection due to the intense focus on the higher status individual (Tukachinsky 2010). However, parasocial conversations also contain substantially fewer swear words compared to romantic relationships, reflecting higher perceived cost of social norm violation due to the relationships larger social distance (Fägersten 2012). Also, consistent with previous findings stating the positive relationship between profanity and social distance (Feldman et al. 2017), swear words appear most commonly for social relationships, followed by romance and family relationships. Finally, the figure also reveals that work- and family-related words match folk expectations: organizational and family relationships are the most likely to use their respective LIWC categories, underscoring the topical differences between relationships. The most frequent five words for each LIWC category (shown in Table 2 in Supplemental Material) confirm that these topical-relationship interactions are not primarily driven by a single word in a category.

Do some relationships talk about more diverse topics than others? Social penetration theory predicts that the variety of topics shared in conversations should increase as relationships further develop (Altman and Taylor 1973). We test this prediction using a topic model to analyze the diversity in communication. Following prior work (Quercia, Capra, and Crowcroft 2012), we measure topical diversity using the entropy of the message’s distribution of topics from a trained LDA model. A 100-topic LDA model is fit using Mallet (McCallum 2002) on a sample of 100K dyads balanced across the five categories and using five tweets per dyad to control for differences in communication frequency. To measure diversity we calculate entropy over the mean topic distribution per dyad, then aggregate by category.

The observed topical diversity (Figure 2) matches predictions from social penetration theory, with more diverse topics (higher entropy) seen in relationship categories that are more likely to contain deeper relationships with stronger ties and to have developed further such as romance, social, and family. In contrast, organizational relationships are less likely to communicate on topics outside their common ground (Marwell and Hage 1970). These results were consistent over several runs with topic models trained on 20 or 50 topics.

Given their different functions, relationships are expected to differ in their network and communication proprieties. We test for these differences by examining the labeled dyads within the larger social network constructed from our entire 10% Twitter sample. Here, two users have an edge if they both mention each other at least once, which results in a network with $\sim$1.1B edges. The dyads with labeled relationships represent a small fraction of this comprehensive social network, as the majority of dyads have not declared a relationship. All the network properties for the dyads in our study are measured according to their users’ statistics in this larger network. Using this network and directed tweets between two people, we consider two aspects of a relationship: (1) communication frequencies and (2) the local network structure around a relationship.

Individuals manage their communications differently according to social relationships or “identities” such as social, work and family (Ozenc and Farnham 2011; Min et al. 2013). By observing diurnal Twitter usage patterns through the timestamps of tweet messages (Golder and Macy 2011), we show the existence of both between- and within-category communication differences across relationship types.

Using our massive volume of communications, we compute a distribution representing the fraction of messages exchanged for each relationship dyad during each hour of a day. With the number of mentions from user u to user v as $\mathbf{{m}}_{u\to w}$, we bin each mentioning tweet according to the hour of the day it was created. We then define $t_{u\to w}\left(i\right)$ as the fraction of mentions produced in the i-th hour so that $\sum_{i=0}^{23}t_{u\to w}\left(i\right)=1$. We restrict our analysis to tweets where a local timezone of the tweeting user is provided along with its global timestamp and convert the tweet to its local time. Also, we only consider cases where the sending user has made at least 5 activities for better smoothing of the distribution.

After we compute the diurnal communication distributions for each dyad, we can aggregate across different relationship categories or subcategories to obtain category-wise diurnal distributions. We provide a comparison of the diurnal distributions aggregated across different categories, shown in Figure 4(a). While all categories share the same pattern of substantially lower communication during dusk and peaks around evening, similar to previous work (Golder and Macy 2011), we can observe slight differences around daytime. To further examine such differences, we center the distributions by subtracting each with a global mean

with $S$ as the set of all dyads across all categories, or in Figures 4(c) and (d), the different subcategories we consider. The centered result (Figure 4(b)) shows a notably higher communication rate for the Organization category during work hours (9-16) which drops afterwards, possibly due to moving away from a work activities and towards friends and family chatter (Farnham and Churchill 2011).

Due to our data scale, we can examine communication patterns within the same relationship category. The trends for the Romance category, shown in Figure 4(c), reveal that even within romance relationships, diurnal patterns can be partitioned into early (“dating” and “boyfriend/girlfriend”) versus stable (“engaged” and “spouse”) stages, where the former communicates more during late hours. One possible reason is that the latter group may consist of more married couples that share the same physical space during the evening and have fewer reasons to communicate through Twitter. Within family relationships (Figure 4(d)), we show that aunt/uncle - niece/nephew communication is more intense during the day compared to parent-child communication. We conjecture that this tendency reflects the lesser degree of perceived closeness between extended families as opposed to direct kin, restricting communication during late hours.

We classify a dyad into one of the five relationship categories on the basis of its behavioral and communication features. The prediction task is conducted with both balanced and imbalanced datasets. Balanced set: Training data uses 200K dyads per category, randomly upsampling organizational relationships which had fewer samples. A random sample of 2K dyads were used for validation data. The test set contained 17,522 dyads per category, where all classes were downsampled to match the least common class, organizational relationships. Imbalanced set: 2M dyads are randomly selected and split by a 8:1:1 ratio into training, testing and validation partitions. In both settings, users were constrained to be in only one partition. We ensure that at least one user of the dyad performs at least five interactions, in order to sufficiently represent a dyad’s interactions. To control for differences in communication frequency, we restrict the data to at most 15 tweets from each user in a dyad, keeping up to 5 tweets for each type of communication: directed tweets, retweets, and public mentions. To avoid data leakage, all tweets that contained the phrase used for labeling the relationship type were removed prior to this process.

To capture information from the different types of communication, we introduce a new deep learning model that performs multi-level encoding to represent these texts. Our base architecture builds upon the model of Huang and Carley (2019) using the parameters of the RoBERTa pre-trained language model (Liu et al. 2019). Each tweet and bio text for a dyad are encoded as constant-length vectors by encoding each with RoBERTa and mean pooling the output layer’s word piece embeddings for the content. These pooled content encodings are fed into a separate 6-layer Transformer network (Vaswani et al. 2017); this second level of encoding summarizes the different sources of text information through its attention mechanism. The output layer of this second-level model is mean pooled as a representation of all communication.

The final dyad representation concatenates the multi-level encoding with (i) character-level embeddings for the username created using 1-dimensional convolutional filters (Kim 2014) and (ii) the four network statistics described in Section 4. This representation is fed through two linear layers using ReLU activation with dropout before a softmax is applied to classify the dyad. RoBERTa models were first pre-trained using 3M training set tweets; then the full classification model was trained end-to-end. Supplemental Material 4.1 provides details of all hyperparameters and training procedures.

Our model can accurately recognize different relationships, attaining a macro F1 of 0.70 on the balanced dataset (Table 3), indicating that relationship categories are identifiable from their network and communication patterns. This performance substantially improves upon that of the XGBoost baseline and random chance. In the balanced setting, the model is most accurate at predicting organizational and parasocial and least at social relationships. We attribute this difference to the intra-class diversity; while organizational and parasocial relationships typically have narrowly-exhibited behavior (e.g., low topic diversity for organizational in Figure 2), social relationships can take many forms, e.g., friends, neighbors. This diversity likely makes the class harder to distinguish.

In the imbalanced setting that reflects the natural distribution of classes, our model offers an even larger performance improvement over baselines. Here, most dyads have a social relationship (76%; cf. Table 3), yet the model is still able to reliably identify all classes. This result highlights the applicability of the model in real-world settings, which is crucial for studying communication dynamics.

Separate ablation studies were performed to test the information from each type of communication (direct, public mention, or retweet) and for the addition of user and network features. Among communication types, public mentions provided the most information (highest performance) for predicting relationship types (0.56 in balanced setting). We speculate that a user has to include context about the mentioned user and their relationship when mentioning them in a public tweet, to provide an explanation to the audience. This information is not required in directed mentions since the expected audience is only the two users. The addition of user features and network features to text features also significantly improves performance, most notably for parasocial relationships, where adding user profiles and network features boosts the F1 score by 0.13. The significantly lower Jaccard coefficients and Adamic-Adar scores of dyads in the parasocial category (Figure 3) likely makes it easier to identify the relationship type, when incorporated as features. Table with full results is in Supplemental Material (Table 3).

As a further validation of the relationship classification model, we test whether the relationships inferred by our model mirror the behavioral properties seen in the labeled relationships. A random sample of 1M dyads is collected where one user has made at least five interactions, mirroring our classifier setup (§5.1). We then collect mention, reply, and retweet activities made between both users of the dyad, and apply the classifiers on these new users. As temporal information is not used in the classifier but shows clear differences by relationship (Figure 4), we test whether the communication patterns in these inferred relationships are similar. The resulting time series from inferred users were highly correlated with those of the labeled data correlations, ranging from 0.928 to 0.947, shown in detail in Supplemental Table 5 and visualized in Supplemental Figure 1. This result indicates the model infers relationships that have highly similar behavior to the labeled data in practice (despite not being trained on these features), e.g., dyads in the organizational category focusing more of their communication during daytime and dropping in volume after work hours for both labeled and inferred data.

As a second test of validity, we examine the distribution of inferred relationships in the random sample. The resulting distribution differs substantially from that of the labeled data, as expected: Social 23%, Romance 23%, Family 14%, Organizational 5%, and Parasocial 36% (cf. Table 2). In the random sample we observe more Parasocial, which aligns with earlier expectations that Twitter is largely a mass media platform rather than a social network (Kwak et al. 2010). However, unlike expectations from earlier work, we observe a significant uptick in stronger social relationships: Social and Romance together account for $\sim$46% of the random sample. We attribute this result, in part, to the requirement that dyads in the random sample must have at least five directed tweets, which likely increases the presence of stronger ties who are more likely to talk more. Our results point to the social nature of Twitter and the need for future work to examine how all relationships are manifested on Twitter—not just those that communicate—to establish to what degree the platform now serves as a social network.

We first select the same number of dyads per relationship category to balance across different relationship types. For each dyad we collect all retweets that occurred between the two users, but remove instances where user mentions occurred in the original tweet. While being mentioned in a tweet is a strong motivator for a retweet to occur (Jenders, Kasneci, and Naumann 2013), here our goal is to understand the interplay between the content of the tweet and the relationship properties of the dyad and thus remove tweets with mentions.

We focus on a balanced prediction task where for each positive retweet that occurred between a dyad, we assign one tweet produced by the same user around the same time which did not get retweeted by the other user. As a result, 50,000 positive and negative tweets were sampled per category and split into training, validation, and test sets at a ratio of 8:1:1.

The models used for the prediction task are also based on a pretrained RoBERTa model for text classification.

Adding relationship information improves performance for retweet prediction, as shown in Table 5, which is known to be a difficult task (Martin et al. 2016). Incorporating relationship types provides a 1% performance increase in F1 score due to an increase in recall by 2%. Here, we show separate performances for test data tweets with and without URLs; tweets with URLs are likely to have different retweet dynamics on the basis of the content in the URL (e.g., retweeting a linked news story versus a personal message) and the different social uses of retweets (boyd, Golder, and Lotan 2010).

Analyzing performance improvement by relationship type reveals a more complex picture of improvement, shown in Figure 5. For tweets containing URLs, the addition of relationship information consistently improves performance, while we observe the largest performance gains for tweets without URLs. In particular, in tweets without URLs, the model sees increases of 2.2%, 3.1% and 2.3% for social, romance, and family categories respectively, signalling that the model can use this social information to decide where a person in that relationship is likely to retweet based on the content. However, the model performs worse for predicting retweets from organizational relations (2.8% decrease) which lowers the overall performance reported in Table 5. In the case of tweets containing URLs, F1 scores increase across all categories, with social (3.8%) and family (4.2%) categories benefiting from the largest gains.

We observe that the larger gains seen in social, romance, and family relationships are due to increased recall (see Table 4 in Supplemental Material). This increase suggests that individuals embedded in social, romance, and family relationships retweet content that is less likely to be retweeted normally (e.g., mundane personal events) because of the nature of the relationship, which the relationship-aware model is able to use to correctly identify the content will be retweeted. This result is further evidence of the interaction between communication patterns and relationship types.