My Tweets Bring All the Traits to the Yard: Predicting Personality and Relational Traits in Online Social Networks

Online Social Networks (OSNs) offer a virtual space in which people connect and interact with others, express themselves, and receive information, in a continuous digital reflection of the real (offline) world. In OSNs, people typically showcase their real self [40] and leave traces in online behavior, which reflect their real-world personality [24]. These traces expose a holistic image of oneself, including both personal characteristics (personality traits) and characteristics that portray their behavior in relation to others (relational traits).

The term personality refers to characteristic combinations or patterns of behaviors, cognitions, and emotional reactions that evolve from biological and environmental factors and form relatively consistent individual differences [13]. The Big Five (BF) or Five Factor model [29] is one of the most distinctive personality theories that constitutes five main traits of human personality representing individual differences in cognition, emotion, and behavior: Openness to Experience, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. On the other hand, relational traits have also been linked with consistencies in social behavior and interaction patterns, with attachment theory [7] as the most emblematic theoretical framework in that respect [31, 43], capturing how individuals experience close relationships to and interactions with others.

Personality traits have been studied in the context of OSNs and the web overall, as findings show that they are strongly linked to OSN use [57], online friendships [60], and online reviews [52]. Moreover, certain prediction models have been proposed [37, 64] to extract users’ psychological background from their online behavioral residue and map it to personality characteristics. However, relational traits such as attachment orientations (AO) have been overlooked in online environments, even though user activity in OSNs heavily relates to social behavior characteristics. This makes the study of a relational profile critical from an application point of view and provides rich information about individuals’ social profile.

The present research aims to address this limitation in OSN research, by studying and predicting both relational traits and personality traits of users. The importance of relational facets of personality for explaining social interaction cannot be overstated. Given that online social media engagement resembles actual social interactions in many respects [15, 30], the need to study how different personality facets are reflected in online expression is particularly compelling. Attachment orientations, a key individual difference of relational orientation, can be derived on the basis of traces found in micro-blogs. Attachment orientations capture one’s notion of the self in relation to others and interpersonal relationships, with attachment theory being one of the key personality theoretical frames to explain actual social behavior [31]. Considering both traditional personality Big Five traits and relational traits is important for (1) providing holistic profiling of OSN users—humans have an integrated profile in which self and social aspects interrelate and affect each other, and joint profiling can be essential for understanding the overall human presence on OSNs; (2) uncovering more traits of people’s psychological and social world has been identified as a direction in OSN research (which currently focuses only on the personality traits) that could help to better explain, analyze, and predict online user behavior [66], e.g., with applications on customer segmentation [46] or digital advertisement environments [17]; and (3) shedding light on social interaction phenomena taking place in OSNs is of great socioeconomic importance, e.g., community formation [32], decision making [42], or information diffusion [12].

To this end, the present article proposes a novel data-driven approach to predict a holistic psychological profile of OSN users, capturing both their personality and relational traits.¹ Building on insights stemming from psychology theory, our approach applies data mining on OSN textual and non-textual data, carefully selects different sets of features for predicting different types of traits, and exploits the inherent correlations in psychological traits, to efficiently predict a complete image of OSN users’ psychological profile. The proposed approach is applied on the Twitter micro-blogging service, which stands as a live, dynamic, and open OSN platform on which people intensively interact, and is largely driven by people’s spontaneous reactions and emotions expressing themselves (personality facet) and interacting with others (relational facet) at the same time.

Specifically, our contributions in detail are as follows:

In this section, we provide an overview of related psychological literature, discuss related work, and highlight several open issues of existing methods. We also highlight the contribution of the present work. Section 3 details the collected dataset and the data mining and feature engineering methodology, and Section 4 presents the design and evaluation of the proposed machine learning predictive model. Finally, we conclude our article and discuss future work in Section 6.

In this section, we discuss psychological personality profiling principles (Section 2.1), personality traces in OSNs (Section 2.2), and personality prediction based on these traces (Section 2.3). We discuss outcomes of the most relevant preceding work, to identify (1) the main factors and techniques that can be used in personality prediction from user traces in online social media and (2) gaps in prior state of the art that motivate the proposed approach.

In this section we present the data collection and analysis methodology we follow, which is used to feed our personality prediction models (Section 4). We first describe the Twitter dataset we collected, labeled with personality scores by the users themselves (Section 3.1). Then, we pre-process the collected data (Section 3.2), extract meaningful features for our analysis and modeling (Section 3.3), and conduct an initial data exploration analysis to characterize the quality of the collected dataset and obtain insights for the design of the prediction model (Section 3.4). Figure 1 depicts an overview of our research methodology.

In this section, we proceed to design a model that predicts the user personality traits from the extracted features (see Figure 1). In the following, we first provide the overview of our approach. We then present and discuss our design choices step by step, along with evaluation results (i.e., prediction accuracy) and comparison with state-of-the-art approaches. Our approach combines expert knowledge from the psychology domain and advanced statistical analysis techniques.

Machine learning approach. The analysis in Section 3.4 reveals that no individual behavioral, emotion, or language metric has significant correlation with any of the individual personality traits. This indicates that simple statistical models (e.g., analytic models or linear models with a few features) are not expected to be able to predict accurately the personality profiles; we experimentally verified this intuition as well. Hence, in our approach we select to use advanced statistical models, i.e., ML models, which have been proved efficient in capturing complex statistical patterns and relations between features and target variables.

Regression, training strategy, and evaluation metric. The scores for the personality traits retrieved from the survey are continuous values in the interval [0,1].² Hence, we select to use regression models that correspondingly predict a (continuous) value for each trait. Before training the models described below, we first split our dataset into 80%-20% test-train. The performance of our model refers to the errors on test sets; however, errors on training sets are also provided.

To measure model performance, we use the RMSE:where \(n\) is the number of samples, \(y_{i}\) is the actual trait value for sample \(i\), and \(\hat{y_{i}}\) is the corresponding prediction of the model. RMSE is the most common metric for regression problems that incorporates the model’s variance and bias.

Base regression model. We tested different regression models (linear, support vector, Gaussian, random forest, etc.) for predicting individual personality traits from the extracted features. The random forest (RF) regressor³ performed consistently better for the majority of the personality traits, and thus is selected as the base regression model in our analysis. We conducted a grid search for the tuning of the hyperparameters of the RF regressors, which led to 100 trees.

Single personality trait prediction and best-performing features (Section 4.1). Using the extracted features, we trained RF regressors to predict each personality trait individually, similarly to previous approaches. However, as discussed earlier we (1) use a larger set of features and (2) provide predictions both for BF and AO traits (the latter have not been considered before in a self-labeled dataset). We find that neither using all the extracted features is optimal for all traits nor was there a single set of features that performs best for all traits. We selected the (sub)set of features for each trait that leads to higher accuracy, built the respective RF regression models, and studied their performance.

Holistic personality prediction with multi-output regression chain models (Section 4.2). As can be seen in Figure 2, and is supported by related literature [28, 50], there are significant correlations between personality traits. Motivated by this observation, we propose a single multi-output model that predicts all personality traits at the same time. The proposed model uses chains of RF regressors (see Figure 3) and exploits the correlations between traits to achieve higher accuracy compared to individual trait predictive models.

Dataset and personality profiling annotation. In order to test our approach in a real-world application, we use a dataset with leaders obtained by Crunchbase (crunchbase.com). This dataset contains business and social media information of organizational employees, including CEOs and staff in executive positions, across organizations throughout the United States. Datasets provided by Crunchbase have been used successfully in recent research [27] in examining the relationship and interactions between leaders and their followers on Twitter. Importantly, as outlined in [27], individuals within the respective dataset who are highly active on Twitter and are followed by others on Twitter are defined as “leaders.” Therefore, on the basis of this definition, even a regular employee (i.e., someone without managerial responsibilities and team oversight in their usual workplace) can be defined as an opinion leader on social media. The case can be made that opinion leaders on social media are indeed leaders due to their social influence on followers’ communication content and patterns and yield large referent and expert power. One example would be a Twitter leader who is known for their expertise in communicating astrophysics and/or whose followers can identify with them [20]. Hence, although these “leaders” can and do use many of the same bases of power as more traditional organizational leaders, they can be regular employees and users. We ranked individuals within this dataset by their Twitter activity and selected a set of 1,063 highly active Twitter accounts (for brevity, we refer to them as “leaders”). For these respective accounts, we extracted tweets posted within the past 6 months from the onset of data collection, resulting in 238,665 tweets. We applied our methodology to extract language, emotional, and behavioral features (see Section 3.3) and fed this information into our pre-trained holistic model, which in turn predicts leaders’ personality traits.

Characterization of the leaders’ psychological profiles. After predicting leaders’ personality scores, we perform a basic statistical analysis to (1) identify the main characteristics of the leader and (2) investigate whether our identified leaders are similar or differ in their characteristics from random Twitter users in our collected dataset.

Table 7 presents basic descriptive statistics of obtained personality scores in the leaders dataset. A more detailed comparison of the distributions is presented in Figure 5 (the x-axis represents each trait value and the y-axis the cumulative distribution function (CDF) of the values). A first observation is that leader personality profiles differ significantly from those of the random users in general; see the very different distributions in Figure 6.

Focusing on individual traits, and concerning the BF traits, we observe that leaders score higher than random users on Openness, the tendency to be open-minded and imaginative. Openness to experience also includes the facet “intellect,” which marks another expected association with individuals who have gained positions of social influence such as opinion leaders or positions of managerial power such as Chief Executive Officers or Heads of Design and similar positions. These results are expected. However, the individuals in our sample scored quite low on Extraversion and Conscientiousness, which is in contrast with the opinion that leaders are dominant and sociable in their environment. This may indicate that behaviors associated with these traits are not easily reflected, and therefore detected, online. As for the AO traits, we observe that leaders’ scores are high on both anxiety and avoidance. A possible interpretation for this is that the leaders in our sample do not wish to maintain successful relationships with other users online and merely use the Twitter platform to advertise their own efforts and accomplishments, as well as to respond to other users who seem to post interesting cognitive (instead of affective) content.

These initial findings highlight an interesting research direction in which our approach could be useful, namely in-depth investigations of possible differences between the real-world and OSN behavior of different groups of people. This may reveal groups or clusters of individuals who tend to have a persona in OSNs very different than their real-world personality, and possibly causing factors for this phenomenon. Clustering users based on their psychological profiles. To obtain further insights on how the two groups differ (random users vs. leaders), we cluster these groups based on the personality profiles (inferred by our model) of the users. Since each personality profile consists of a high-dimensional vector, we employ a t-distributed stochastic neighbor embedding (t-SNE) graph that reduces dimensionality by mapping profiles on a two-dimensional space. We present the t-SNE visualization in Figure 7, where users are represented by dots (red for random users, green for leaders) and user coordinates correspond to their psychological profile (as mapped in the two-dimensional space). Two dots that are close to each other denote that the users have similar profiles, since t-SNE calculates a similarity measure based on the distance between points. It can be clearly seen in Figure 7 that our holistic model is able to separate different types of users in distinct clusters; i.e., all users in each cluster are of the same role (color), and the users would be indistinguishable only if the colors in the same cluster were mixed.⁶ This is an interesting finding, which indicates that our holistic psychological profile model (and the seven dimensional profiles it predicts) can be used toward efficiently clustering/classifying different groups of Twitter users.

Moreover, to understand the role of AO, BF, and their combination on the clustering task, we evaluated the clustering results for three different cases: HM (AO+BF chain model), IM (AO+BF independent model), and BF independent model. While Figure 7 shows that there exist clearly four clusters, for a more complete analysis, in order to select the optimal number of clusters for each task we utilized the elbow method [34], and to evaluate the clustering results we use the silhouette coefficient, which calculates how similar the objects on a group are compared to other groups. Elbow method analysis in Figure 6 shows that the optimal number of clusters is among three and four and as a result we use these numbers as a proxy for the optimal number of clusters. We also account for k = 2 as the baseline. Table 8 presents the different silhouette scores for each of the three cases. It can be seen that when taking into account the AO information (either in the HM or the IM models), the clustering improves (i.e., higher silhouette value) compared to the model that uses only the BF characteristics, and this finding holds for all number of clusters we tested, indicating the robustness of using both AO and BF traits.

While these preliminary results demonstrate the usefulness of our work in this direction (a more detailed analysis would be beyond the scope of this article), we believe that this can motivate further research in the topic, e.g., by considering different groups of users (such as artists, politicians, influencers, etc.) or classification techniques.

Anxious and avoidant attachment orientations are important constructs in how people express themselves in interpersonal relationships [43]. In combination with BF personality traits, these relational traits have been associated with several offline and online environments that require individuals’ presence and/or interaction, for example, consumer behavior [45], social media use [57], social network friendships [60], as well as the formation of online communities. These associations highlight the important role of personality traits, both in self-expressions and as relational traits, in interactions with others. However, to the best of our knowledge, no previous studies have provided prediction algorithms to detect relational traits based on self-report measures. The present article addresses this gap by providing a machine learning detection and prediction approach with regard to both relational and Big Five personality traits in one model, providing a holistic image of the social self.

We find that the strong inter-correlations between personality and relational traits indicate that employing a multi-output model for personality prediction exploiting the best predictors among language, emotion, and behavioral residue combined better represents the psychological mechanisms behind personality expression and relationship interactions online, achieving at the same time higher accuracy performance than other state-of-the-art models.

The holistic psychological profile can find applicability in a number of studies and fields. We provided an initial demonstration through a use case by applying the model on Twitter users that are organizational leaders in the real world. The initial findings show groups are clustered efficiently based on similar personality characteristics. Similar approaches can be applied to different groups of persons and contribute in social science studies, investigation of the “pretend persona” phenomenon [5] (i.e., deviations between the offline and online behavior of people, similarly to those we found in the leaders dataset), user classification in OSNs, design of recommendation algorithms, information diffusion, and so forth. We believe that these are interesting directions for future research. Furthermore, the proposed approach could be combined with other psychological models reflecting individual differences in order to uncover more psychological patterns reflected on our behavior in OSNs applied in different case-specific samples creating a personality map. As additional future research we propose that the investigation of personality and attachment profiling in the social networking context by combining information of users in a network could reveal the mechanism by which people tend to form relationships online based on how they do it in real life. Finally, the study of information diffusion in personality networks could reveal which personality type is able to propagate certain kinds of information and emotions more effectively with applications on the marketing and news sector.

We present a detailed analysis and results on the correlations between features and traits in our dataset.

Several associations were found between anxious attachment and language features. Anxious individuals tended to use language that conveys authenticity, indicating the extent that the language used is personal and self-revealing [51] (\(\rho = 0.15\)), and feelings (\(\rho\) = 0.17; words such as “feels,” “touch,” etc.), as well as religion (\(\rho\) = 0.15; words such as “church,” “altar,” etc.). In addition, individuals who scored high on anxious attachment were also more likely to use filler words (\(\rho\) = 0.17, words such as “I mean” or “You know”). Avoidant attachment, on the other hand, was significantly and positively associated with authenticity (\(\rho\) = 0.14), mentions of health-related words (\(\rho = 0.13\); words such as “clinic,” “flu,” “pill,” etc.), and space (\(\rho\) = 0.13; words such as “down,” “in,” “thin,” etc.). We also found that avoidant individuals were less likely to mention leisure-related words (\(\rho = -0.13\); words such as “cook,” “chat,” “movie,” etc.). Avoidant persons also seemed to use comparisons (\(\rho = 0.14\); words such as “greater,” “best,” etc.) and numbers frequently (\(\rho\) = 0.18; words such as “second,” “thousand,” etc.), which hints at prior work that shows avoidant individuals tend to have a negative view of others, and hence would be more likely to see themselves in a more positive light. With regard to behavioral features, it seems both anxious and avoidant AOs were significantly and negatively related to the number of user mentions (\(\rho = -0.20\) and \(\rho = -0.17\), respectively). Hence, individuals who scored higher on either of these traits tend not to directly refer to others within a social media context.

Regarding the BF dimensions, we found several correlations among language and topic use, emotions, and behavioral patterns within each trait. For example, agreeableness was significantly and positively correlated with the expression of joy (\(\rho\) = 0.14), and extraversion was negatively correlated with fear (\(\rho = -0.14\)). We also observed that openness was significantly and negatively correlated with surprise (\(\rho = -0.20\)), while highly neurotic individuals were more likely to express sadness (\(\rho = -0.17\)). In addition, conscientiousness was negatively correlated with retweeting behavior (\(\rho = -0.16\)) as well as tweet length (\(\rho = -0.16\)), while conscientious individuals were more likely to pick lengthy screen names (\(\rho = 0.13\)). Moreover, we noticed that highly neurotic individuals were more likely to have more followers (\(\rho\) = 0.18) and tweet other people’s content on a regular basis (\(\rho\) = 0.14).