Sam Naghshineh

Analysis of the Tie Strengths in Social Networking Using Face Recognition

Social media is currently not suitable for differentiation between friends and acquaintances in the list of contacts. In reality, people’s relationship fall in a wide range of strength and it has been a research topic for many decades now. This variance when put in the scenario of the social networks causes issues with privacy and the level of information that can be shared with newly added contacts. There has been some research done in that area that tried to apply the concepts of tie strength to social networking environment. In this paper we are trying to look at the problem from the different perspective and to answer the question of whether the analysis of the pictures via face recognition from the social events can help to bridge the gap between understanding the relation between people more precisely and later use that finding in social networking web-sites. The model is based on more than 1000 pictures from different events and analysis of the relation between the selected individuals at the events. It showed a good performance and relatively low error rate. We have implemented the face recognition algorithms that helped us collect the statistical data on each person at the event and come up with a visual representation of their connection and relation to other people using weighted graphs and adjacency matrices.

Keywords

Social media, social networks, face recognition, relationship modeling, ties, sns, tie strength, photo sharing, event photography, php, sql, imagemagick

Introduction

One said “Friends are the roses of life: pick them carefully and avoid the thorns” and another said “Everyone is a friend, until they prove otherwise”. Linking both of these proverbs to the current situation with social networking web sites and the notion of friends lists, it is possible to summarize the problem and we are dealing with right now. How many of those multiple hundred-word long friends lists are filled with genuine friends. But then, who decides who is ‘genuine’? Should we add everyone on Facebook for the sake of networking and socialization, or only certain selected individuals? Then how to decide what information should be allowed to be viewed and by who? What if it is not only friendship, but also a job reference, what kind of information is allowed to stay on the profile and what should be hidden? Can the process of this decision-making be automated?

In the era of online social networking and sites like Facebook, it has become a recent trend to try to understand the relationship between people, the strength of their ties to compound a more sophisticated and robust network that would not only allow communication between the members but would also understand between which members this communication is more likely to happen.  Social media is based on the relationship between people, but that relationship is different from person to person. Going through the list of friends of Facebook one cannot name everyone to be a friend, but rather an acquaintance, classmate or colleague. Such connections are useful for information sharing, finding a job, or simply networking, but would probably be not so appropriate when sharing some personal information. Currently, social media does not differentiate between the stranger and a friend, letting users decide whether they would like to accept particular person’s friend request and possibly limit their profile. Unlike most social networking websites, which indicate in a binary format whether two people know one another, we can hypothesize not only whether they know one another but also how well do they know one another from the tie strength results

There have been many research papers written on this subject. The most recent is by Eric Gilbert, where he examines the tie strength of the contacts on Facebook. In his paper Gilbert addresses the question of whether “social media data predict tie strength” [1]. He uses the method of involving the participants who were willing to share their Facebook date and automated data collection. Gilbert talks about how their research on tie strength can improve the social networking design elements, as well as other aspects such as a better privacy control, “messaging routing, friend introduction and information prioritization” [1].  However, it is very important to realize that their paper investigates the tie strength inside the social network itself, whereas we are taking a completely different approach and dealing with the event-based photography, though looking for the similar results in the end.

In our research we tried to find out whether it is possible to identify the tie strength between the people based on the collection of data from the event-based photography. In this effort we are going to try to evaluate the level of people’s interaction based on the pictures collected from the events. Such images can be taken with cell phones, digital camcorders and, certainly, digital cameras. An event such as wedding, party, family reunion, conference, or any other of the kind usually has numerous photographers, whether they are amateur or professional, and hence several hundred photos are likely to be taken. Now, after collecting all these pictures at the centralized location, and having this enormous data on hand, we thought of whether it is possible to be able to use it to identify the patterns of social interaction. More specifically, the main question that we are going to investigate in this paper – is it possible to measure consistency and magnitude of interaction and social behavior by using face-recognition mechanisms on the pictures from the events? Furthermore by analyzing the accuracy of the image recognition we can conclude whether to use this method as a fallback to image tagging. Expanding beyond that, we will also try to see if it is possible to identify the kindred or family connections.

We begin by first introducing the definitions of tie strength and principles that fall behind it. We then try to analyze the previous research and what lessons we can take out of it when dealing with our own research. Then we introduce the detailed method of data collection, data analysis, and data visualization. Finally, we will make the conclusions that we came to, and will wrap up with the possible limitations and errors of our research and future work that can possibly improve our experiment.

The strength of the ties

In this section we would like to review the notion of the tie strength and what it means for our research in particular. We are not going to go into the deep analysis of the predictive variables as Gilbert does in his research, but we are going to get a general overview of the notion of weak and strong ties and the difference between the two. Mark Granovetter first did the introduction of the notion of the tie strength in 1973. His paper became a base for all the future research on the tie strength in social networking environment. Tie strength is a “combination of the amount of time, the emotional intensity, the intimacy (mutual confiding) and reciprocal services which characterize the tie”. According to the author, “weak ties are less likely to be socially involved with one another than are our close friends strong ties”[2]. Gilbert expanded on that definition a little, and suggested that strong ties are the connections or “people you really trust, people whose social circles tightly overlap with your own.” [1]. A lot of time those are your friends, people of the same age, interests, who you might know for many years, and who share certain life activities. “The young, highly educated and the metropolitan tend to have diverse networks of strong ties”. [1, 3] Conversely, weak ties are the ones between merely acquaintances, providing the information on the social status, job, publications, work connections, but not of the personal matter. [1] Therefore the network of the people that have weak ties is of low-density (“one in which many of the possible relational lines are absent”), and strong ties – densely knit connections with close friends. [2]

The subject of investigating the tie strength has become really hot in the last decades, due to the increase of the popularity of social networking web sites and networking in general. There has been numerous research works done on this subject. However, the proportion of researchers who uses tie-strength is overwhelmingly larger than the number of empirical studies that have made an attempt to measure tie-strength [6, 4]. In those of them that did, authors try to identify the criteria for measuring the tie strength. Granovetter proposed amount of time, intimacy, intensity, and reciprocal services. Researches that worked on this subject later expanded that and included such factors as network topology and social circles. [5]. In general, there can be a lot of different factors that can relate to the strength of the ties between the individuals, therefore Gilbert selected few most important ones, such as communication reciprocity [7], possessing at least one mutual friend [8], recency of communication [9], and interaction frequency [10, 2]. Gilbert talks about the work of Marsden and the use of the survey data from three metropolitan areas to precisely unpack the predictors of tie strength [1, 3]. The limitation of the Marsden’s work was the survey itself that asked to pick the three closest friends given some characteristics of friendship.

In our research we are, in a way, fixing that limitation and proposing a new method of measuring the tie strength. We are going to talk about the method in greater details in the Method section, but in general our work primarily differs by the way we trying to collect the data about the people, with the goal in mind to be more fair and objective, and do not depend on the subjective opinions of people of what is friendship and close relationship. We are deviating from all the proposed criteria of measuring the tie strength and rather proposing our own, which is the number of appearances a person has with other people at the pictures from different events.

CONTRIBUTION

The main goal of our project was to design and develop a web-based photo sharing service that would be used for the event photo collection, organization, and analysis of the question of tie strength between the people at the events. Our web site has the functionalities that all current photo-sharing web sites offer and more. We believe that what we came up with does not currently exist and has a great value to the community of photographers and the field of the social networking research and development.

RELATED WORK

We partially mentioned some related work that was done in the field, however there is definitely more to add. The topic of tie strength has been into the spotlight recently and not only in regards of the online social networking but also at the communities and groups. There are around 766,000 research articles that mention the notion of the tie strength (based on Google Scholar search). A lot of researches tried to tie this notion with studying individuals and organizations. [2]. There is a wide range of applications of that research. There were studies that strong ties improve mental health and are good for a balanced life [11]

We found the paper by Gilbert to be really closely related to our research. He talks about how their research on tie strength can improve the social networking design elements, as well as other aspects such as a better privacy control, “messaging routing, friend introduction and information prioritization” [1]. However, it is very important to realize that his paper investigates the tie strength inside the social network itself, whereas we are taking a completely different approach and dealing with the event-based photography, though looking for the similar results in the end. The paper goes in too much of details about the different types of variables of the tie strength. In our research we are deviating from such a thorough analysis of the notion of tie strength itself, but rather looking for the possibility to determine using face-recognition if there are any ties at all between the people at the event. Nonetheless, the thoroughness of the research by Gilbert gives us a lot of insight on what we can possibly do with our data collected and somewhat enhance the results that were collected by him and his team. Finally, the authors talk about their research being limited due to the fact that they only analyzed the connections of the people on Facebook. In a way, we are expanding that research by analyzing the tie strengths between people that came to a social event and possibly providing a deeper insight on the subject.

METHOD

Web Data Collection

In order to facilitate the project requirements of being able to establish a connection between social behavior and event photography there comes a rise in demand in increasing the sample size to accommodate a well distributed group of test subjects (event pictures and albums). For the ease of collection, project continuity, and future expansion we have created a PHP based website for image collection and manipulation. This website feature serves as a share point for collection of event photography for the users such that they can collaborate among themselves as anonymous individuals by uploading as many pictures from every person’s camera into the system. The user begins such process by first registering and setting up the shared album in which will serve as the picture collection pool. The link of the shared pool can then be sent to other members or can be preset by the event so that all members are aware of the pool. Evidently if/once the website gains popularity – like Facebook and other social media – the users will automatically seek the shared album and share their pictures. One may ask what might attract the users to such application, and the answer is that the website provides automatic image manipulation (orientation, duplicate removal, duplicate angles etc.) tools, which will be applied to the album thread once enough pictures (either by time lapse or number of image uploads) has been uploaded (indicated by the user during album creation). Once the thread is closed the website begins to apply image tools to pictures and notify the album contributors that the album pictures are now organized and compressed, ready to download. This image collaboration allows every person that contributed to the album to have a complete set of the entire event’s pictures.

Further to the website, this technique to attract users and utilize the masses to upload can be used to drive images into the website. Although this sort of implementation is in the works this method of image collection can be utilized to gather large amounts of images for further analysis. Thus utilizing the website we have gathered a large pool of images from several events and will be using further added tools such as image recognition on top of image manipulation to analyze social behavior based on event photography.

The Setup

The above implementation was achieved using a dedicated Quad core server that can withstand high server load to image processing using imagemagick and its C API namely magickwand. The website was designed ground-up using a combination of PHP web scripting language and SQL database for image database. The snapshot of the web site is present at Figure 1. Images are collected using Java Script to allow large uploads (quantity and size) and are preprocessed via GD2[1] image manipulation library for thumbnail processing and are in parallel manipulated via imagemagick for duplicate management and orientation handling. Once the filtration process is completed the face recognition process begins. Although not entirely meshed the image recognition API has been borrowed from face.com with use under the Python language the faces are recognized in command prompt and an output is generated giving a temporary name to the recognized face combined with some EXIF[1] (Exchangeable image format) data for every scanned image. Due to the limited number of resources the full implementation and meshing of the face.com recognition was not completely meshed with the website for automatic face recognition and training (face recognition calibration). As a results the images and their resulting output has been gathered manually for data analysis purposes. Due to the large amounts of data by increasing the scope and size of the sample number of people selected for social behavioral analysis, for every event no more than twelve randomly nominated individuals who appeared in multiple events were selected for analysis. The reason for this size is that the problem becomes exponentially larger since for every person added to sample size that person must be analyzed against every other person within the group, thus increasing the amount of work load. Thus due to the manual labor the analysis is applied to a smaller group.

Figure 1 – MyEventExchange website (link)

Figure 2 – Sample picture from one of the events after running Face Recognition Algorithm

Figure 2 shows an example of the image that was outputted after running face recognition algorithms on it. In addition to people being identified, there is other supporting data that was found. For example, the woman on the picture is not obviously smiling, and the engine identified that saying that her smiling is in the range of 3%. On the other hand, the men are smiling, and the algorithm was able to detect that with 60% and 61% accuracy. Neither of them is wearing glasses, and therefore attribute for glasses is false with a high percentage of confidence. On some other pictures, where people were wearing glasses the algorithm was successful in identifying that. It is also very accurate in gender identification and the positions of their heads. This data is not necessarily useful in our research in particular, because we are mainly dealing with face recognition itself and linking that to the relationship between people. However, additional data can always be used in other application and we are going to talk about this briefly in the concluding section of this paper.

Data Analysis

Following the raw output of image recognition results from the pool of images uploaded to the website we can then begin analyzing the social behavior from the number of appearances people have had with one another. Several assumptions and deduced theorems were utilized in conveying the results. These include that given the resources (images only) if a person is socially active within an event they are more likely to appear in photos. Although the person taking pictures is more likely to have pictures of themselves on their own cameras we have extended this lemma such that a person is socially active if they appear within a lot of photos containing as many people as possible. In other words the score given to the person at hand is comprised of the total normalized number of images that other people share with this person. The normalization is done by taking number of pictures any pair has had and dividing that by the total number pictures the group of people who are being analyzed are tagged in. This method of analyzing every person against another person will yield the tie strength of the social behavior within that event. Unlike most social networking websites which indicate in a binary format whether two people know one another, we can hypothesize not only whether they know one another but also how well do they know one another from the tie strength results. By obtaining the number of images any pair share and dividing that by the total number of tagged images the person at study has had and normalizing that factor, we can obtain a percentage of how well two people know one another.

Figure 3 The weighted graph of the group of people from the bigger events – demonsrates a more sparse distribution

Figure 4 Example of the weighted graph that is more densely knit

To generate the tie strength graph we are utilizing a biograph to present the data. The biograph in this case is unidirectional since we are assuming (although not necessarily true) that person A knows person B just as much as B knows A. This assumption is critical since bidirectional data analysis is not feasible due to limited resources (images can only give unidirectional results). In order to convert the raw data and analyze it in this mezzanine core between raw and biograph, the data is placed within an adjacency matrix. An adjacency matrix is a way of representing which vertices (or nodes) of a graph are adjacent to which other vertices. However, normally, within the matrix the data are binary, but in our case our data varies from 0~1 in terms of percentage thus not only representing which nodes are connected using which vertices but also how strong the vertices are. Better yet how well is one node connected to another node?

Figure 5 Example of an adjacency matrix used in the data analysis

From the matrix in Figure 5, the first notable factor is that the diagonal of the matrix is 0. This clearly represents the assumption that a person is acquainted with themselves (although a person should technically know themselves 100% however for the sake of mathematical normalization we are assuming 0).

Using this method of data analysis we can then organize the entire sample size of participants into this matrix, which will represent their social measured behavior for the all events that they have attended. Since the sample size is deemed rather small compared to the number of attendees for every event the same procedure is done for entirely different group. Thus there are two independent groups that do not have any sort of acquaintance that have attended three events together. Of each group 10~11 random individuals are selected for sampling whom have attended most of the events. Finally, for each of the independent groups an adjacency matrix is generated to mathematically map their measured social behavior for use for data visualization.

Once the data analysis has been revealed the same process is repeated except this time a survey is conducted on each person to have him or her provide us with how much they feel they know a person. These sets of values can serve as a reference point to compare and contrast with the accuracy of measuring tie strengths from photos compared to real results.

Figure 6 Normalized data from the face-recognition based on the hundred-people events

Figure 7 Actual Data Collected from the Interviews

Figure 8 The difference of the actual data with the data collected from the face recognition

Figure 10 Data after the face recognition of the people from the events of smaller capacity

Figure 11 Actual data collected from the interviews

Figure 12 The difference of the actual data and the data from face-recognition to show the error rate

Data Visualization

From the adjacency matrices we were able to create the weighted graphs using Matlab software. In our case the graphs are unidirectional, because the number of appearances of person A with person B is the same as person B with person A.  This part was the most critical in our research, because it allowed determining what type of connection people have between each other, which is the whole objective of this research. Currently, social networking websites can only determine whether there is a connection between certain people (LinkedIn), but they do not know the strength of that connection. The weighted graphs that we constructed show exactly how related people are.

Figure 3 and Figure 4 shows two different examples of the possible outcomes. Figure 3 is a visualization of the connections among the group of people that were randomly selected at the events of the capacity of hundred people and more. We looked at the number of pictures that they appear together based on the three different events. The graph is rather sparse due to the fact that the probability of the two people having their picture taken multiple times is smaller than between the two people at the smaller event. The graph shows the connections between the 11 people with the weights assigned to the sides. From this graph we can conclude that Sam Naghshineh knows Amirali Rahnamoon the best and those who have the zero relation know each other the least or don’t know each other at all. Looking at the Figure 6 we can verify the same results. From Figure 4, it is clear that the events that we collected pictures from were much smaller, hence the probability of the two people taking picture together increased, and, as a result, we got a denser graph. It also makes sense because the smaller the event the more likely it is for people to know each other.

There is also a big chunk of group pictures from these events that added some number to the total pictures where all the examined people are tagged. Looking at the Figure 4 we can conclude that Logan knows Kyumo and Eva knows Paul the best. Numerical results can be seen from the Figure 10. The following results are going to be proven to be close to the reality by the interviewing the selected people from the group. Those results are presented in the next section.

Interviewing

We were able to conduct the interviews with the people that we picked to analyze for the tie strength relation. There was one question that we asked: How strong is your relationship with this person? (Answer based on a scale from 0 to 1)We then were able to construct the adjacency matrices for the answers that people provided in order to be able to compare the results with our actual face-recognition experiment. The data results are present in Figures 7 and 11.

RESULTS

We were able to conclude that the results of our experiments are successful and the model works well. In order to determine the accuracy of our experiment we did few analyses.

Bar Graph Comparisons

In order to determine the accuracy of our experiment, we decided to graph the averages of the data collected from the experiments and the interviews and compare the trends with each other. Figure 13 represents the comparison that we did among the group of people that we picked from the bigger events. The red bars are visualizing the actual data and the blue graphs – experimental. From this graph we can conclude that the trend of each set of data is the same, in other words, the curves of each data sets would be similar. There is, certainly, some difference in numerical values for every person; however, we think that this difference makes sense. First of all, people tend to exaggerate and overestimate, hence when we asked them to identify the friendship the numerical value of the friendship, or the strength of the relations with other people, it is quite normal to have the results that might be not exactly the same as experimental data showed. On the other hand, the difference is also coming from the number of pictures taken at the event. Every person did not take pictures with other people every time there was a picture taken. They might have been missing, talking to someone else, or simply not in the mood for taking the pictures. We should take into account this human factor, and instead of comparing the actual numbers, look at the trends that the bar graphs show. In the end, they are very similar, and hence, our method works quite accurately.

Figure 13 Comparison of the averages of actual data (red bars) with experimental data (blue bars) of the Group 1

Now, looking at Figure 14, we can notice that the bar graphs do not show exactly the same behavior. The data in this chart is from the second experiment, when we collected information on people who attended smaller events. The error is mostly present in the right side of the chart, where people said that they do not actually know those people who they were on the pictures with.

However, the fact, that the events were smaller, the probability of taking group pictures, where everyone is present, or pictures of people who might not necessarily know each other too well is higher, therefore we see the difference in actual and experimental averages.

Figure 14 Comparison of the average values of experimental versus actual data of the Group 2

Figure 15 “Butterfly” graph showing the distribution of the expected data (pink area) and actual data (blue dots)

The better understanding might be obtained after looking at Figure 15, that represents the data that we expected to see (pink area) and the experimental data distribution (blue dots). The four areas in that graph represent four different scenarios that we considered: High Association among people and Low Appearance on the pictures together, High Association – High Appearance, Low Association – High Appearance, and Low Association – Low Appearance. Intuitively it makes sense to see more data for High Association – High Appearance and High Association – Low Appearance, because people who know each other well tend to take more pictures together. It is also expected to see some of the pictures in the other two categories, but, certainly, their amount should be significantly smaller. Now, those are of Low Association – High Appearances is what bringing the error to the graph in Figure 14. At the smaller events, people who did not know each other well, ended up taking a lot of pictures together, due to the fact that a lot of group pictures were taken. This is definitely something to consider, and keep in mind when dealing with the future iterations of this research.

Error calculation

In order to determine the error rate for our experiment, we could not use the standard formula because some of the connections had zeros, due to the fact that people did not know each other or simply did not pictures together, and we could not divide by zero. Therefore what we are doing to figure out the error rate is subtracting the experimental results from the actual results collected based on the interviewing people. The results of this can be seen in Figure 8 and Figure 12 for the two groups of people respectively. We decided to set up the threshold of the error being significant to be a 30%. Based on this, we determined that the model incorrectly identified 6 out of 55 connections in case of the bigger events (Figure 8), which is 11%, and 8 out of 55 in case of the smaller events (Figure 12), which is 14.5%. We think that this demonstrates that the model was pretty accurate and it is indeed possible to use face recognition to determine the relations between the people at the events. The method works best on people who have more pictures taken at the events, because it is easier to determine how strong the ties are due to a better calibration of their social behavior via pictures. With people who took fewer pictures, it is rather less accurate because there might be some limitations on why there is a smaller amount of pictures taken. It might be the issue of people being not social, and taking pictures only because they do not know others well (in which case our model is accurate). However, it might also be the case of people being not in the right place and not in the right time. With the case of the smaller group the error might come from the fact that some people, who don’t know each other end up being in the group picture, and the face recognition captures that, so without even knowing each other too well, people end up being tagged together (we showed that in the precious section). Furthermore, those who weren’t involved in too many pictures had a smaller sample size thus it is harder to pinpoint their social behavior within an event, as a result higher error rate is witnessed for those who didn’t appear in too many pictures.

LESSONS AND CONCLUSIONS

We ran the experiment of face recognition on the selected group of individuals, and obtained the experimental and actual data that we were able to compare and decide on the success and accuracy of the experiment. We conclude that our model is accurate and can be used for answering the set research question of whether it is possible to identify the tie strength between people at the events based on the photography face recognition.

There are several aspects that we could deduce out of this research. Firstly, the image recognition method works best within a group of pictures with a sample selected group of individuals that have appeared nominally within pictures. We already discussed that the opposite can cause higher rate of error.

The method that we used in this paper can be improved either. As long as we used mostly the manual labor, once an automated implementation is applied computational power can be used to increase the sample size to as large as possible, which will definitely add the level of precision and accuracy for the selected sample. Additionally, we neglected many assumptions that would make this research too complicated and non-uniform to make any conclusions. For example, it is possible that the strength between the people who take picture together is much higher than between the people who just appear on the group picture. On the other hand, when two people take picture together, it might not necessarily mean that they are best friends, but rather colleagues with weak tie of relationship. Just looking at the pictures, it is rather hard to make such conclusions; therefore we were assuming uniform weight distribution for every connection. However, this is, certainly, something to keep in mind for future research.

The method of face recognition can be used in social networking web sites as a substitute for manual picture tagging, and can save a lot of time. When tagging, people can already be granted the necessary level of information access that would be allowed for them to view at the profile of another person, since the tie strength between the two are going to be determined.

This research can also be expanded to a lot more areas than just social networking. For instance, it can be a very useful tool for the analysis among the community of photographers, as each picture has EXIF data provided by their cameras, so why not collect that data and analyze the settings, the quality of the pictures, and the professionalism of the photographers.

The portal itself might become a new centralized location for people to share their pictures, which is going to be a more convenient and easy solution given the wide range of photo-sharing web sites and all of their disadvantages.

Additionally, due to the fact that the data analyzed from face recognition is more than just a name of a person, but rather whether a person is smiling, wearing glasses, male/female, we can possibly use that data in other applications. For example, when looking at the pictures, and counting the number of smiling people, we can possibly identify whether the event that they attended was fun or enjoyable for them, and if not the organizers can think how to improve it in the future. Other application is to see how many people are there who look similar, to possibly identify family connections or just to see who look alike. Some of the other suggested questions based on the data obtained might be: Is there a way to suggest people to invite to a party based on this? How do people strengthen their relationships during events?

Concluding, this topic has great opportunities for further research and might be applied to various interesting applications.

REFERENCES

1. Gilber, E. Predicting Tie Strength with Social Media.Proceedings of the 27^th International Conference on Human Factors in Computing Systems. ACM, New York, USA 2009
2. Granovetter, M. 1983. The Strength of Weak Ties: A Network Theory Revisited. Sociological Theory, 1, 201-233.
3. Marsden, P. V. 1981. Core Discussion Networks of Americans. American Sociological Review, 52(1), 122– 131
4. Petrócz, A. Measuring tie-strength in virtual social networks. Connections,27(2): 39-52
5. Burt, R.  Structural Holes: The Social Structure of Competition. Harvard University Press, 1995.
6. Mathews KM, White MC, Soper B, & von Bergen CW. 1998.Association of indicators and predictors of tie-strength. Psychological Reports, 83: 1459-1469.
7. Friedkin, N. E. 1980. A Test of Structural Features of Granovetter’s Strength of Weak Ties Theory.  Social Networks, 2, 411–422
8. Shi, X., Adamic, L. A., et al. 2007. Networks of Strong Ties.  Physica A: Statistical Mechanics and its Applications, 378(1), 33–47
9. Lin, N., Dayton, P. W., et al. 1978. Analyzing the Instrumental Use of Relations in the Context of Social Structure.  Sociological Methods Research, 7(2), 149-166
10. Gilbert, E., Karahalios, K., et al. The Network in the Garden: An Empirical Analysis of Social Media in Rural Life. Proc. CHI, 2008. 1603–1612
11. Schaefer, C., Coyne, J. C., et al. 1990. The Health-related Functions of Social Support. Journal of Behavioral Medicine, 4(4), 381–406