- Open Access
Coevolution of a multilayer node-aligned network whose layers represent different social relations
© The Author(s) 2017
Received: 22 February 2017
Accepted: 13 October 2017
Published: 6 November 2017
We examine the coevolution of three-layer node-aligned network of university students. The first layer is defined by nominations based on perceived prominence collected from repeated surveys during the first four semesters; the second is a behavioral layer representing actual students’ interactions based on records of mobile calls and text messages; while the third is a behavioral layer representing potential face-to-face interactions suggested by bluetooth collocations.
We address four interrelated questions. First, we ask whether the formation or dissolution of a link in one of the layers precedes or succeeds the formation or dissolution of the corresponding link in another layer (temporal dependencies). Second, we explore the causes of observed temporal dependencies between the layers. For those temporal dependencies that are confirmed, we measure the predictive capability of such dependencies. Third, we observe the progress towards nominations and the stages that lead to them. Finally, we examine whether the differences in dissolution rates of symmetric (undirected) versus asymmetric (directed) links co-exist in all layers.
We find strong patterns of reciprocal temporal dependencies between the layers. In particular, the creation of an edge in either behavioral layer generally precedes the formation of a corresponding edge in the nomination layer. Conversely, the decay of a link in the nomination layer generally precedes a decline in the intensity of communication and collocation. Finally, nodes connected by asymmetric nomination edges have lower overall communication and collocation volumes and more asymmetric communication flows than the nodes linked by symmetric edges.
We find that creation and dissolution of cognitively salient contacts have temporal dependencies with communication and collocation behavior.
In this paper we use a dataset containing longitudinal information on a group of individuals in a multilayer node-aligned network to examine dependencies across different types of relations (for details on data collection see ). We use this rich source of information, hereafter referred to as the NetSense data, to build three distinct network layers linking individuals over time. One is a nomination layer constructed from subjective reports of significant contacts, another is a behavioral layer constructed from electronic records of communications, and the third is a behavioral layer, constructed from bluetooth records indicating spatial collocation of students.
In a previous version of this paper presented at the conference , we analyzed dependencies between the first two layers. Here, we aim to understand the relationship between all three layers to shed the light on the link between nominations, communication behavior, and spatial propinquity. This is important, since, as we note in the following section, the exact relationship between these forms of human connectivity is a long-standing, but understudied, problem in social network analysis [3, 4].
Motivating social science background for the research
Social scientists have traditionally distinguished various dimensions of human connectivity [5–9]. Perhaps the most well-studied dimension of this type in fields like anthropology and sociology is cognitive saliency. This can be defined as the subjective prominence of a given contact for an individual at a particular point in time [6, 10, 11]. Empirically, cognitive salience can be measured as the likelihood that an individual will “nominate” another individual as an important contact with regard to a given relation (e.g., friend, advisor, discussion partner, frequent interlocutor). The classical method used to study this dimension of connectivity among individuals is the network survey [6, 7, 12, 13], in which individuals are presented with a “name generator” (to elicit some predetermined number of salient nominees from memory) and a “name interpreter” (to collect relevant information on each nominee). This approach has generated a great deal of knowledge (usually at the level of “ego-networks”), about those contacts who are subjectively the most important to each individual. A key advantage of the NetSense data is the availability of such periodic network surveys recording each participant’s most cognitively salient contacts.
A related dimension of human connectivity is frequency of interaction [7, 9, 14]. An important finding of network analysis in sociology is that, while frequency and cognitive salience usually go together , the correlation between the two factors is much weaker than would be expected if these two dimensions were two indicators of the same underlying construct such as “tie strength” [7, 9]. Instead, research has established that persons can have high rates of communicate interaction (e.g., established either via observational or self-report methods) with contacts who were not cognitive salient or considered particularly close [5, 7, 8]. In the same way, some non-negligible proportion of the most cognitively salient contacts may be characterized by relatively low rates of communicative frequency . Overall, however, the question of whether frequency of interaction precedes, and therefore leads to, cognitive salience or whether salience leads to more interaction remains a highly debated issue [9, 14–16].
Finally, both classic and more recent work in social networks points to spatial contiguity or propinquity as an important indicator of social connectivity . Just like with the relationship between cognitive salience and frequency of interaction, recent work shows that spatial contiguity is an independent dimension of human association, since it can vary independently of the other two factors mentioned. Persons may spent a lot of time (e.g., in workplaces, classrooms, and so on) in spatial proximity to people who are not particularly salient to them or with whom they seldom interact directly . However, research also shows that strong social ties tend to emerge among people who spend time together in the same physical location , such that both cognitive salience and frequency of interaction may end up being the result of external factors that channel people into the same physical spaces. This is what  has referred to as “proximity” mechanisms governing the formation and maintenance of dyadic connections among individuals.
Research questions and hypotheses
Social network researchers have proposed studied dependencies between all three of the aforementioned dimensions of human association. Most of this work however, has looked at the connection between two of the aforementioned three dimensions at a time [14, 16], usually with an eye towards establishing the causal precedence of one factor (e.g., frequency of interaction or propinquity) over the other (e.g., the cognitive salience of a given contact). Seldom, however, has the mutual connection between these three factors been explored systematically. We use the rich set of data collected as part of the NetSense project to do just that.
In what follows, we begin by establishing the validity of the cognitive salience and interaction frequency measures that were collected as part of the NetSense data. We thus begin by exploring the correlations between layers in a multilayer network having three types of links, one based on interaction frequency, another based on collocation frequency, while the third is based on cognitive salience. Social network theory leads us to conjecture that these two dimensions should be positively correlated, with cognitive salient ties displaying (on average) more frequent behavioral interaction than non-salient ties. In essence, we should expect that persons connected in the nomination layer, should have higher interaction frequencies [3, 5, 7, 9].
We then address the question of precedence between interaction frequency and cognitive salience. Drawing on models that see behavior as preceding cognition, we hypothesize that nodes linked by communication edges with large weights at a given point in time should be more likely to appear as cognitively salient contacts in the future. Following , we test the same hypothesis with regard to collocation behavior: We thus expect that the more an individual encounters another person in a proximate physical location at a given point in time, the more likely it is that person will appear as a cognitively salient contact in the future. A key issue is which behavioral mechanism, frequency of interaction or collocation, is more important in determining future salience in the nomination [9, 16]. This is an issue that has not been investigated in previous work but that we tackle here using state-of-the-art machine-learning methods.
While the hypothesis that behavior in large part determines cognitive salience is plausible, recent work on the culture-network link also proposes that the reverse arrow of causation is also equally plausible: That going from cognitive salience to behavioral interactions [17, 18]. We propose that a dissolution event in the nomination layer (e.g., being mentioned as a significant contact at time t but then not mentioning that contact at time \(t+1)\) should be a accompanied by decreasing levels of behavioral interaction in terms of both communication and collocation events. Essentially, the disappearance of a nomination link should lead to a gradual decay of intensity of the communication and collocation layers.
We use the temporal features of the NetSense data to examine the question whether variation in cognitive salience determines the communication and collocation interactions. We take advantage of the differences in duration among cognitively salient ties (differentiating between newly salient and long-standing ties) hypothesizing that nodes connected by long-standing nomination ties should exhibit higher levels of communication and collocation than the ones linked by newly emerging ties.
In addition, we go beyond establishing differences in behavioral profiles of ties based on their levels of cognitive salience, and examine the question of whether the behavioral temporal signature of social ties differs depending on whether their nodes are linked in the nomination layer or not. We hypothesize that cognitive salient ties will also exhibit behavioral activity in days of the week and times of the day that are associated with informal sociability in this population (e.g., weekends and evenings), and that this feature differentiates them from ties whose nodes are disconnected in the nomination layer.
Finally, we aim to examine, for the first time, coupled asymmetries and non-reciprocities along the nomination and behavioral layers. In traditional networks built from subjective reports of cognitive salience, a common phenomenon is asymmetry. This refers to the situation where contact B is cognitively salient for person A, but not the reverse : A is not mentioned by B as a cognitive salient contact. Work in social network analysis shows that this situation is fairly common in human social networks, even for ties defined by friendship and intimacy . Here we examine if asymmetries in nominations are connected to non-reciprocities in the two-way flow of communication in the behavioral layer. Recent work also shows that non-reciprocity is more common in human social interactions that would be expected from traditional social and anthropological theory pointing to the “norm” of reciprocity . We hypothesize that these two phenomena are empirically linked: nodes connected by asymmetric nomination edges should be connected by behavioral edges characterized by non-reciprocity. We also connect the phenomenon of cognitive asymmetry with that of behavioral tie decay , and hypothesize that asymmetric cognitive ties should exhibit steeper rates of decay in the behavioral communication layer.
Related work on multilayer network dynamics
A model to generate two social networks synthetically, with both networks coevolving, capturing the properties of both networks is introduced in . A rapidly evolving network based on games is studied in . Nodes in this network have varying incentives to build links. We observe similar behavior in the NetSense data, where certain edges have incentive to develop into an edge in one of the networks, while others do not. A wide swath of previous work, impossible to fully review in this limited space, in social network analysis has investigated dependence and evolution of connections across multiple networks (for a review see ).
More generally, multilayer networks have been studied extensively in the Network Science literature. They have been formally defined in  which also includes a discussion of related topics such as multilayer networks, multilayer node-aligned networks, multiplex networks, and hypergraphs. The NetSense network can be classified as node-aligned multilayer network because it has one set of nodes and three types of edges connecting these nodes. This reference discusses several properties of multilayer networks and describes how link prediction is done across multilayer heterogeneous networks in [26, 27]. This reference defines also a “meta-path” approach, where sequences of different relations are used as features for link prediction. Novel community detection approaches for evolving multi-slice networks have been discussed in  and they will be useful for future work related to our paper.
A study similar in design to that of NetSense is discussed in . The authors analyze a multilayer network of high school students, including a face-to-face interaction layer and a friendship survey layer. They study the difference in structural properties of three layers in their successive paper . Smieszek et al. study a multilayer network between conference participants in  including a proximity sensor interaction layer, and the self-reported friendship layer. The authors study in what ways these two layers differ. In contrast to these two papers, our study focuses on the evolution of the edges in our multilayer network. The dataset that contains only records of communication between people is presented in . The participants are high school students transitioning to college. The authors study the change of communication volume between pairs of students as they come into contact with new people, to discover the evolution of communication edges in response to external events. In our study, we focus on how changes of edges in one layer impact edges in other layers. Stopczynski et al. introduce in  the Copenhagen dataset defining a multilayer network that includes communication layer, a face-to-face interaction layer, and a Wi-Fi signal-based proximity layer. Several papers on the Copenhagen study describe how different social structures, like groups congregating at the same locations, are discovered using different layers of the network. Large-scale communication record datasets typically focus on basic features, such as call patterns, temporal features of communication records, and evolving communities in such network, as exemplified by [32, 33]. Miritello et al.  study how individuals adopt different tie formation strategies to activate and maintain ties. The Copenhagen dataset and the large-scale communication record dataset contain only behavioral edges, while an important element of our study is to investigate how behavioral edges impact changes in the nomination layer.
Additions to our previous work
The present research builds on, but moves beyond, our previous work, which explored how two-layered social network coevolves in time. This work was published in the Proceedings of Fifth International Workshop on Complex Networks and their Application . Here, we introduce a second behavioral network layer, based on the bluetooth collocations of nodes. We demonstrate how the behavioral collocation layer coevolves with the nomination layer (based on nominations) as well as the first communication-based behavioral layer. We then look at the different behavioral stages in the evolution of the networks. In addition, we also included the analysis of temporal dependencies between the two behavioral layers and the nomination layers. Social ties and temporal features have been studied in . However, here we study the temporal dependencies between different kinds of ties and the coevolution of these ties over time.
NetSense data and the networks
In this section, we introduce the NetSense data  and the networks derived from it. The data were collected at the University of Notre Dame. At the start the Fall semester in 2011, 200 incoming freshmen were enrolled in the NetSense study. Over 150 participated until their graduation in the Spring of 2015. Students participating in the study received free smartphones with unlimited voice and text plans as an incentive for participation. We obtained time-stamped logs of communication records for all study participants. These data contain information on the date, time, and duration (for calls) and character length (for text messages). Data for the first four semesters (lasting from the Fall of 2011 to the Spring of 2013) of the project were available for this study.
In addition to this, we have information about the bluetooth interactions between study participants. The bluetooth interactions are time-stamped. However, the bluetooth interaction data are available only from October 2011 to May 2012. From all the bluetooth interactions, we filter the interactions which are most likely to be face-to-face. Every bluetooth interaction has an associated signal strength value, which is referred to as the received signal strength indicator, or RSSI. The RSSI values vary between 0 and − 120 dB, a value close to 0 implying high quality of the signal. RSSI values above − 65 dB are generally inferred to be most likely face-to-face as mentioned in . Therefore, we consider only the interactions where the RSSI values are above − 65 dB.
Students participating in the NetSense study list up to twenty nominees at the beginning of each semester. Students were asked to list the names of those people with whom they thought they spent the most time communicating or interacting. Recent work has shown that this type of name generator is the most likely to produce an unbiased list of significant contacts . Below, we refer to these cognitively salient contacts as nominees. For each student, nominees could be inside or outside the NetSense study. Because students were asked to also provide the primary phone number of each significant contact, we can link each of those mentioned in the survey to the time-stamped smartphone data.
A summary of the NetSense layers
Frequency of evolution
Nominations in surveys
Calls and texts
Methodology and results
Dynamic coevolution of the nomination, collocation, and communication layers
In this paper, we are interested in understanding coevolving dynamics between the communication and collocation connectivity layers (indexing behavior) and the layer of nominations (indexing cognitive salience of contacts). We also seek initial validation of these coevolution dynamics. Given that the period of time for which NetSense has gathered information covers the initial stages of a college cohort, we can expect that students will nominate others as significant contacts, communicate, and spend time in proximity of one another within a relatively closed environment. We postulate the existence of multiple relationships between the coevolving collocation, communication, and nomination layers.
We hypothesize that there is a delay between the onset of changing behavior, observed either in communication or collocation, and the nomination of another individual as a cognitively salient contact. Communication (mutual calls or texts) or collocation (i.e., spending time together, whether in a class or in a club or just casual meet-ups) would transition in a later time period to a contact appearing as significant enough to be nominated in the survey. We study how cognitive salience is affected by behavioral interaction whether in terms of communication or collocation. In this section we see how behavior can be used to predict the formation and persistence of cognitive salient relationships. We also expect to see the converse occur, where a decline in cognitive salience leads to a diminishing of communication and collocation.
Do nodes connected in the nomination layer have stronger edges in the behavioral layers?
We explore the differences in communication and collocation volumes between nodes who are and who are not connected in the nomination layer. Communication volume is measured by the number of calls and messages exchanged between dyads in a semester, while collocation intensity is measured by the number of bluetooth proximity detection events between the same dyads in a given semester.
Difference between nodes connected and disconnected in the nomination layer in terms of weight of their communication and collocation edges
Do higher weight behavioral edges turn into new nomination edges?
We explore how dynamic changes in the cognitive salience of contacts, as indexed by the formation of new nomination links, correlate with the volume of interaction along the corresponding link in the communication and collocation layer. We examine whether an increase in the cognitive salience of a given contact, is preceded by higher rates of behavioral communication and collocation. We also examine whether links that decline in cognitive salience, as indexed by a dissolution of the corresponding nomination edge, are characterized by lower volumes of communication and collocation the semester before.
Difference in communication volume between pair of nodes who are to-be-nominees and the ones who are not-to-be-nominees
Predicting the formation of links in the nomination layer from the strength of edges in the behavioral layers
Here we examine whether formation and existence of edges in the nomination layer can be predicted from their respective edge weights in the communication and collocation layer. We take the weight of the behavioral edge, defined by the communication or collocation volume, and consider whether behavioral edges above a threshold of the behavior volume can predict whether we observe a formation event in the nomination layer. We expect that edges below the threshold should be more likely not form an edge in the nomination layer in the future.
Prediction of nominated contact based on the volume of communication in the communication behavioral layer
Prediction of future-nominated contact formation based on volume of communication between the corresponding nodes in the particular semester
In Fig. 3a, b we show how the performance of prediction changes with changing thresholds. This also reflects upon the differences between the distribution of behavioral weights of nominees and non-nominees and between the to-be-nominees and the not-to-be-nominees.
Difference in recall values when different combinations of behavior are used
Collocation + communication
Difference between persistent and dissolving nomination edges in terms of weights of the corresponding communication behavioral edges
Difference between persistent and dissolving nomination edges in terms of the weights of their corresponding collocation behavioral edges
Do newly formed nomination edges differ from older nomination edges in terms of their communication and collocation volume?
Here we examine whether there are systematic differences between newly formed and older edges in the nomination layer, in terms of how strongly connected they are in the communication and collocation layers. To this end, we measure and compute the difference in the volume of communication and collocation between nodes joined by older (at least one semester old) and newly formed edges in the nomination layer.
Difference in behavioral communication volumes between nodes connected by the old and new nomination edges
Newly observed nominations
Nomination contact older than one semester
Difference in the numbers of behavioral collocations between nodes connected by the old and new nomination edges
Newly observed nominations
Nominations older than one semester
We also observe that as these newly formed nomination edges age, the nodes connected by them come to have communication volumes similar to, or perhaps slightly higher, than nodes connected by edges in the nomination layer that have existed for a longer time. To shed further light on this issue, we examine communication volumes between nodes connected by the nomination layer in the 3rd and the 4th semesters, and we divide them into edges which were created in the 2nd and the 3rd semesters, respectively, and edges which existed since the 1st semester. We call the former moderately old edges and the latter very old edges. We observe that the nodes connected by the moderately old edges carry on an average of 49 calls and 903 calls, while the nodes connected by the very old edges exchange 29 calls and 795 messages. We infer that communication between nodes that are connected in the nomination edge increases gradually, but then finally stabilizes over a period of time.
Is link decay in the nomination layer followed by a weakening of the corresponding edge in the behavioral layers?
In this section we examine the question of whether behavioral links tend to weaken and dissolve after the corresponding edges in the nomination layer decay. To that end, we measure the rate at which pairs of nodes that become disconnected in the nomination layer also become disconnected in the communication and collocation layers. Then we compare that with the rate at which behavioral links dissolve at random. We want to ascertain whether nodes that first experience a dissolution event in the nomination layer are more likely to dissolve edges in the behavioral layers than a random dyad does.
To do this, define the quantity BDND as the average link dissolution rate in the behavioral layers for persons who are not connected in the nomination layer, and BDNC as the average behavioral link dissolution rate for persons that are connected in the nomination layer. In the third and fourth semesters, BDND is significantly greater than BDNC, while the reverse is observed in the second semester. We observe values of 64, 55, and 50% for BDND for the three semesters, and 42, 74, and 62% for BDNC. We also measure the rate at which the nodes connected by edges in the nomination layer that persist into the following semester dissolve their behavioral edges, and denote it as BDNP. We find that BDNP is always 0, meaning that if the nomination link persists then so does the corresponding behavioral link. Yet, we rarely observe dissolution of collocation after the contact is no longer nominated to the list of significant contacts. This implies that people may continue to frequent the same places even after they are no longer cognitively salient contacts.
Are the temporal collocation signatures of nodes connected in the nomination layer different from those who are not connected?
In this section we examine the question of whether students who are connected in the nomination layer exhibit distinct patterns of temporal collocation behavior. In particular, we ask whether cognitive salient contacts are more likely to be in the same place during days of the week and times of the day more likely to be associated with sociability (such as weekends and evenings, respectively) than students who do not share a link in the nomination layer. We also look at whether temporal communication and collocation behavior differs for individuals who will become connected in the nomination layer in the future. A positive result here would indicate that we can use observed changes in temporal behavioral signatures among dyads to predict increases in the cognitive salience of a given contact over time.
To do this, we look at different temporal features of the collocation behavior and explore differences among four groups of dyads: nominees (dyads connected in the nomination layer), non-nominees (dyads disconnected in the nomination layer), to-be-nominees (dyads who will form a link in the nomination layer in the future), and not-to-be-nominees (dyads that will in the nomination layer). One key behavioral feature is the number of bluetooth collocations on weekdays versus the number of collocations on weekends. We include Friday evenings into the weekend count. Another is the number of collocations that happen during weekday days versus the number of collocations that happen on weekday evenings and nights. We define collocations that happen in the day, as the ones that happen between 8 a.m. and 6 p.m. and that evening collocations happen after 6 p.m. and before 8 a.m. We observe the behavior differences between nominees and non-nominees and between to-be-nominees and not-to-be-nominees and see if they are similar. Similar behaviors might indicate latent expression of nomination edges.
Differences between nominees and non-nominees in numbers of collocations between weekdays and weekends
Differences between nominees and non-nominees in numbers of collocations between evenings and daytimes
Differences between to-be-nominees and not-to-be-nominees in the numbers of collocations between weekdays and weekends
Differences between to-be-nominees and not-to-be-nominees in the numbers of collocations between evenings and daytimes
Emergence of cognitively salient contacts through defined stages of interaction
In this section, we examine the dynamic coevolution of the two behavioral layers with respect to the nomination layer. We believe that while both behaviors, communication, and collocation are fairly interchangeable when it comes to prediction of contact formation in the nomination layer, they are not necessarily synchronized and one might precede the other. First, we determine whether collocation is a predictor of future increases in communication, or whether communication predicts increases in future collocation. We then study whether the behavioral factor that follows the other is an intermediary mechanism linking the casually pre-eminent behavioral factor (e.g., collocation or communication) to edge formation in the nomination layer.
Emergence of cognitively salient contacts: from collocation to communication to cognitive salience
To look at the evolution of the two behavioral layers, we consider the paths from collocation to significant contact formation occurring through a communication relationship. We would like to know if the communication relationship is an intermediary variable between mere collocation and the formation of link in the nomination layer. We explore whether the formation of a communication edge is correlated with the weight of the corresponding collocation edge. We also examine whether forming or increasing the strength of an collocation edge precedes a corresponding edge creation in the communication layer.
Difference in the numbers of collocations between communicators and non-communicators
Difference in the number of collocations between nodes to-be-communicators and not-to-be-communicators
Step-wise evolution of significant contact formation
After confirming that increasing weight in the communication layer is very likely an intermediate stage preceding the formation of ties in the nomination, we can observe how some edges in the nomination layer emerge gradually over time, from mere collocation to communication finally leading to nomination as cognitively salient contact. In the first semester, we find that students who start communicating in the second semester, but who do not achieve enough cognitive salience to nominate one another in the network survey, have on an average 74.6 collocations, while those who would not communicate have on an average 25.9 collocations. Now out of these communicators in the second semester, the ones who move on to become nominees in the third semester, have on an average 166.2 collocations, while those that do not become nominees have on an average 55.1 collocations in the second semester. So we observe how some of the emergence of nomination links progress from collocation to communication, gradually increasing the number of occasions in which they find themselves in the same place.
Are changes in collocation behavior more strongly affected by the dissolution of communication or nomination edges?
In this section, we examine the question of whether collocation behavior is affected more strongly by edge dissolution in the nomination layer or edge decay in the communication layer. To do this, we measure how many nodes with edges continue having collocation after the dissolution of communication, and if this behavior differs from collocation after the dissolution of links in the nomination layer. We observe that 42% of nodes initially connected by a communication link continue to have bluetooth collocations after this edge is dissolved. On the other hand, 93% of dyads continue to experience collocation after a contact is no longer nominated. This implies that communication behavior is more important in determining whether people spend time in the same place than cognitive salience. Conversely, this means that people spend time in the same place with persons that are not cognitively salient to them.
Analysis of asymmetric nomination edges in relation to the behavioral layers
As established in previous work in social network analysis , network links premised on cognitive salience, such as the edges in the NetSense nomination layer, have the property of potentially being asymmetric: one person may nominate the other as a cognitively salient contact but the other may fail to reciprocate (A nominates B but B does not nominate A). This is contrast to communication edges, which have continuous weights and in which relations can be more or less reciprocal but almost never completely asymmetric . Behavioral collocations, by definition, have to be symmetric (if A is in the same place as B, then B is in the same place as A) .
Are there differences in communication and collocation volumes between nodes connected by the asymmetric and symmetric edges in the nomination layer?
Difference in communication volumes between nodes connected by asymmetric and symmetric nomination edges
Difference in the numbers of collocations between nodes connected by asymmetric and symmetric nomination edges
Are nodes with asymmetric edges in the nomination layer more likely to be connected by non-reciprocal communication edges?
Next, we examine whether nodes connected by non-reciprocal communication edges are also more likely to have asymmetric communication patterns. We define a non-reciprocal communication edge as one in which one node initiates communications with the other node more often than the reverse . We compare communication imbalance between nodes connected by asymmetric and non-symmetric nomination links. To measure non-reciprocity in the communication layer, we first compute the ratio of the volume of communication in which the source node is the initiator to the volume of communication in which the destination node is the initiator: we call this quantity One-Sided Communication Factor, OSCF. We multiply the number of calls by 10, since messages are about 10 times more frequent than calls and add the product to the number of messages. Using this result, we measure the percentage of non-reciprocated communication for nodes connected by both asymmetric and symmetric in the nomination layer.
Difference between symmetric and asymmetric edges in the fractions of having non-reciprocal communication, OSCF
Asymmetric edges (%)
Symmetric edges (%)
Is behavioral edge dissolution faster for nodes linked by asymmetric edges in the nomination layer?
We examine if nodes connected by asymmetric nomination edges are more likely to experience behavioral edge decay than nodes connected by symmetric nominations. To do so, we measure the survival probabilities of behavioral edges between nodes connected by asymmetric and symmetric edges across all semesters. We find that nodes connected by asymmetric nominations are significantly more likely to experience decay and dissolution in the communication layer than the nodes connected by symmetric edges in the nomination layer in all three semesters.
Nodes joined by asymmetric edges in the nomination layer have dissolution probabilities of 90, 87.5, and 50% in the communication layer in each of the three semesters. In contrast, the corresponding probabilities for dyads joined by symmetric nomination links are 72, 66, and 16% in each of the three semesters. We also observe an overall downward trend in the dissolution probability in the communication layer. Initially, these are very high for the first semester, but they decline steadily over time. However, even in the third semester, nodes connected by asymmetric nomination edges are more than three times more likely to dissolve their behavioral edges than nodes joined by symmetric nominations.
Communication behavior profile: the “non-reciprocal sender” profile
We classify nodes that are more likely to be involved in non-reciprocal communication as non-reciprocal senders. We then examine the communication behavior profile of these nodes to see if the non-reciprocal sender profile differs from reciprocal sender profile. The goal is to verify if nodes with different communication profiles are more likely to experience their changes in the nomination layer, given the well-known psychological aversion to lack of reciprocity [3, 19, 37]. We find support for the hypothesis that non-reciprocal senders have a larger churn of nominees, in the observation that non-reciprocal senders retain 7, 16, and 38% of their links in the nomination layer, while reciprocal senders retain 25, 50, and 88% of their nomination links in the succeeding semesters.
Implications of the results
A fundamental question in the analysis of social networks concerns itself with dependencies between links based on cognitive salience, such as those elicited from traditional network surveys, and behavioral linkages indicative of direct communicative exchange, or providing the potential for the formation of close relationships such frequenting the same physical location [3, 12, 37, 38]. Dependencies between cognitive salience, propinquity, and behavioral interactions have been difficult to study in the past, mainly due to lack of availability of dynamic, ecologically valid data in which the temporal dependencies of network ties across these different layers could be examined [9, 15].
In this paper, we have provided a unified empirical treatment of the temporal coevolution of network layers capturing these three key types of human connectivity. Our results provide confirmation for classical lines of network theory [6, 37, 38], while revealing novel insights about the linkage between cognition, communication, and behavior. First, we show that there are systematic dependencies between the cognitive salience of contacts and communication and collocation behavior. All else equal, persons tend to communicate and spend time in the same place as those contacts that are cognitively salient to them. This is particularly the case for those contacts whose cognitive salience persists over time. This is consistent with the idea that cognitive structure is a key determinant of behavioral structure , and that frequency of interaction and time-spent together are important components of the concept of tie strength . However, our results also show that just in the same way that cognition is predictive of behavior, behavior is predictive of dynamic changes in the relative salience of contacts. Using state-of-the-art machine-learning techniques we showed that future changes in the cognitive salience of contacts can be predicted from pre-existing (and potential) interactions, both in terms of communication and propinquity behavior.
Previous work on the cognitive salience of social contacts using network surveys has shown that such ties are subject to what has been referred as decay; this is the phenomenon whereby a person may nominate another as a cognitively salient contact at time t but fail to do so at time \(t+1\). While the phenomenon of tie decay is well studied in social network analysis , it has not been previously linked to coevolution dynamics of behavioral ties. In this paper we showed that there are systematic links between tie-decay dynamics in terms of cognitive salience with respect to changes in dynamic behavioral layers. For ties that decline in cognitive salience over time, some changes in communication behavior and very minimal changes in collocation behavior occur. This implies that changes in cognitive salience are less predictive of behavioral changes than the reverse. Nevertheless, we do find that cognitive salientties exhibit systematic differences in terms of the temporal profile in which behaviors are enacted. Cognitively salient ties tend to be activated (either via communication or collocation) at days of the week or times of the day much more likely to be indicative of informal sociability. One implication of this novel result is that this behavioral temporal signature could be used to predict cognitive salience in the absence of subjective information of social ties.
In examining the linkage between these three different forms of human connectivity, an empirically validated model of the emergence of cognitive salient contacts suggests itself. According to our findings, collocation behavior emerges first, which leads to increases in communication behavior, which then leads to a contact rising up high enough in the cognitive salience hierarchy to be mentioned as a significant tie in a network survey. The process model suggested by our results is consistent with classic work on the evolution of social contacts from psychology and sociology [19, 37, 41]. We are able to provide systematic empirical evidence for the first time here.
Finally, our empirical work speaks to the fundamental role of asymmetry and non-reciprocity in human connectivity. Previous work has pointed to the fact that the cognitive salience relation can be asymmetric: A can be salient to B but B may not consider A salient . In the same way, previous work has shown that communications can be either reciprocal or non-reciprocal . Our empirical work connects these two lines of research for the time, showing that asymmetry at the level of cognitive salience is connected to non-reciprocity in communication behavior in systematic and intuitive ways. All else being equal, asymmetric nominations lead to non-reciprocities in communication. Not only that, asymmetry at the level of cognitive salience predicts tie decay in the communication layer, while also predicting the churning of ties at the level of the individual.
Limitations and suggestions for future research
The present work advances theory and research in social network analysis, especially with regard to the dynamic coevolution of social ties across multiple connectivity layers. However, our results also open up a variety of questions that cannot be answered given the limitations of the data, and which should be the subject of future work. In particular, moving our framework to a larger population beyond the college student setting, and ascertaining whether our results hold in other human interaction foci (such as work organizations) is important. Linking cognitive salience to other subjective features of social relations, such as emotional closeness, role-relations (e.g., friendship versus kin ties), the exchange of resources (e.g., advice, or emotional support), and looking at how these edge-level variables in the nomination layer interact with communication and collocation behavior also seems like a pertinent subject of future research. Finally, while only suggestive at this stage, expanding the work on asymmetries and non-reciprocities in both cognitive salience and behavioral interactions seems like a promising avenue of future research. This work could look at the individual, dyadic, and contextual correlates of reciprocities and non-reciprocities in interaction as these interact with subjective symmetries and asymmetries in cognitive salience. In all, the work reported here opens up multiple avenues for future work at the intersection of cognition, behavior, and human social networks.
This work is the result of a joint effort where all authors contributed equally to defining and shaping the problem definition, methodology, deciphering results, implementation, and manuscript. BA, as the first author, took the lead in composing the first draft of the manuscript, while everyone else edited it. All authors read and approved the final manuscript.
Acknowledgements and funding
This work was supported in part by the Army Research Laboratory under Cooperative Agreement Number W911NF-09-2-0053 (the Network Science CTA) and by the Office of Naval Research (ONR) Grant No. N00014-15-1-2640. The views and conclusions contained in this document are those of the authors.
The authors declare that they have no competing interests.
Availability of data and materials
The data are owned by University of Notre Dame and administered by Prof. Omar Lizardo who will share the data for legitimate research purposes under the modest requirement of acknowledging the source of data.
Consent to publish
All authors have approved the manuscript and have given the consent to publish.
Ethics approval and consent to participate
The Institutional Review Board (IRB) has reviewed the NetSense study and has approved it; the IRB Number is FWA 00002462.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
- Striegel A, Liu S, Meng L, Poellabauer C, Hachen D, Lizardo O. Lessons learned from the netsense smartphone study. In: Proceedings of the 5th ACM workshop on HotPlanet (HotPlanet ’13). New York: ACM; 2013. p. 51–6.Google Scholar
- Bahulkar A, Szymanski BK, Chan K, Lizardo O. Coevolution of two networks representing different social relations in NetSense. In: Proceedings of the international workshop on complex networks and their applications, studies in computational intelligence series 693. Berlin: Springer; 2016. p. 423–34.Google Scholar
- Hammer M. Implications of behavioral and cognitive reciprocity in social network data. Soc Netw. 1985;7(2):189–201.View ArticleGoogle Scholar
- Lazer D. The co-evolution of individual and network. J Math Sociol. 2001;25: 69–108.View ArticleMATHGoogle Scholar
- Campbell KE, Lee BA. Name generators in surveys of personal networks. Soc Netw. 1991;13(3):203–21.View ArticleGoogle Scholar
- Hammer M. Explorations into the meaning of social network interview data. Soc Netw. 1984;6(4):341–71.View ArticleGoogle Scholar
- Marsden PV, Campbell KE. Reflections on conceptualizing and measuring tie strength. Soc Forces. 2012;91(1):17–23.View ArticleGoogle Scholar
- Roberts SG, Dunbar RI. Communication in social networks: effects of kinship, network size, and emotional closeness. Pers Relationsh. 2011;18(3):439–52.View ArticleGoogle Scholar
- Rivera MT, Soderstrom SB, Uzzi B. Dynamics of dyads in social networks: assortative, relational, and proximity mechanisms. Ann Rev Sociol. 2010;36:91–115.View ArticleGoogle Scholar
- Brewer DD. Forgetting in the recall-based elicitation of personal and social networks. Soc Netw. 2000;22(1):29–43.View ArticleGoogle Scholar
- Marin A. Are respondents more likely to list alters with certain characteristics? Implications for name generator data. Soc Netw. 2004;26:289–307.View ArticleGoogle Scholar
- Feld SL, Carter WC. Detecting measurement bias in respondent reports of personal networks. Soc Netw. 2002;24(4):365–83.View ArticleGoogle Scholar
- Merluzzi J, Burt RS. How many names are enough? Identifying network effects with the least set of listed contacts. Soc Netw. 2013;35(3):331–7.View ArticleGoogle Scholar
- Grabowicz PA, Ramasco JJ, Gonçalves B, Eguíluz VM. Entangling mobility and interactions in social media. PLoS ONE. 2014;9(3):e92196.View ArticleGoogle Scholar
- Kitts JA. Beyond networks in structural theories of exchange: promises from computational social science. In: Advances in group processes. 2014. p. 263–98.Google Scholar
- Wiese J, Kelley PG, Cranor LF, Dabbish L, Hong JI, Zimmerman J. Are you close with me? are you nearby? Investigating social groups, closeness, and willingness to share. In: Proceedings of the 13th international conference on Ubiquitous computing. New York: ACM; 2011. p. 197–206.Google Scholar
- Fuhse JA. The meaning structure of social networks. Sociol Theory. 2009;27(1):51–73.View ArticleGoogle Scholar
- Vaisey S, Lizardo O. Can cultural worldviews influence network composition? Soc Forces. 2010;88(4):1595–618.View ArticleGoogle Scholar
- Carley KM, Krackhardt D. Cognitive inconsistencies and non-symmetric friendship. Soc Netw. 1996;18(1):1–27.View ArticleGoogle Scholar
- Wang C, Lizardo O, Hachen D, Strathman A, Toroczkai Z, Chawla NV. A dyadic reciprocity index for repeated interaction networks. Netw Sci. 2013;1(1):31–48.View ArticleGoogle Scholar
- Raeder T, Lizardo O, Hachen D, Chawla NV. Predictors of short-term decay of cell phone contacts in a large scale communication network. Soc Netw. 2011;33(4):245–57.View ArticleGoogle Scholar
- Zheleva E, Sharara H, Getoor L. Coevolution of social and affiliation networks. In: Proceedings of the 15th ACM SIGKDD international conference on knowledge discovery and data mining. New York: ACM; 2009. p. 1007–16.Google Scholar
- Skyrms B, Pemantle R. A dynamic model of social network formation. In: Adaptive networks. 2009. p. 231–51.Google Scholar
- Pattison P, Robins G, Wang P, Snijders TAB, Koskinen J. The co-evolution of multiple networks. In: Sunbelt XXVI international social networks conference vancouver. 2006.Google Scholar
- Kivela M, Arenas A, Barthelemy M, Gleeson JP, Moreno Y, Porter MA. Multilayer networks. J Complex Netw. 2014;2(4):203–71. doi:10.1093/comnet/cnu016.View ArticleGoogle Scholar
- Zhang JKX, Yu PS. Transferring heterogeneous links across location-based social networks. In: Proceedings of the 7th ACM international conference on web search and data mining. New York: ACM; 2014. p. 303–12.Google Scholar
- Zhang J, Yu PS, Zhou Z. Meta-path based multi-network collective link prediction. In: Proceedings of the 20th ACM SIGKDD international conference on knowledge discovery and data mining. New York: ACM; 2014. p. 1286–95.Google Scholar
- Mucha PJ, Richardson T, Macon K, Porter MA, Onnela JP. Community structure in time-dependent, multiscale, and multiplex networks. Science. 2010;328(5980):876–8.View ArticleMATHMathSciNetGoogle Scholar
- Smieszek T. Contact diaries versus wearable proximity sensors in measuring contact patterns at a conference: method comparison and participants’ attitudes. BMC Infect Dis. 2016;16:341.View ArticleGoogle Scholar
- Mastrandrea R, Fournet J, Barrat A. Contact patterns in a high school: a comparison between data collected using wearable sensors, friend diaries and friendship surveys. PLoS ONE. 2015;10:e0136497.View ArticleGoogle Scholar
- Saramaki J, Leichtb E, Lopezb E, Robertsc S, Reed-Tsochasb F, Dunbare R. Persistence of social signatures in human communication. Proc Natl Acad Sci. 2014;111:942–7.View ArticleGoogle Scholar
- Stopczynski A, Sekara V, Sapiezynski P, Cuttone A, Madsen MM, Larsen J, Lehmann S. Measuring large-scale social networks with high resolution. PLoS ONE. 2014;9:e95978.View ArticleGoogle Scholar
- Saramaki J, Moro E. From seconds to months: multi-scale dynamics of mobile telephone calls. Eur Phys J. 2015;88:164. doi:10.1140/epjb/e2015-60106-6 View ArticleGoogle Scholar
- Miritello G, Lara R, Cebrian M, Moro E. Limited communication capacity unveils strategies for human interaction. Sci Rep. 2013;3:1950.View ArticleGoogle Scholar
- Miritello G. Temporal patterns of communication in social networks. Berlin: Springer Theses. 2013. p. 294–6.Google Scholar
- Liu S, Jiang Y, Striegel A. Face-to-face proximity estimation using bluetooth on smartphones. IEEE Trans Mob Comput. 2014;13(4):811–23.View ArticleGoogle Scholar
- Davis JA. Statistical analysis of pair relationships: symmetry, subjective consistency and reciprocity. Sociometry. 1968:102–19.Google Scholar
- Carley K. Group stability: a socio-cognitive approach. Adv Group Process. 1990;7:1–44.Google Scholar
- Marsden PV, Campbell KE. Measuring tie strength. Soc Forces. 1984;63(2):482–501.View ArticleGoogle Scholar
- Burt RS. Decay functions. Soc Netw. 2000;22(1):1–28.View ArticleMathSciNetGoogle Scholar
- Newcomb TM. An approach to the study of communicative acts. Psychol Rev. 1953;60(6):393.View ArticleGoogle Scholar