This chapter summarizes some of the discussion among presenters and other participants throughout the workshop. Among the many points raised and questions asked of presenters, this summary focuses on those aspects of the discussion relevant to the purpose of this workshop, as articulated in Chapter 1: to gather information for the Decadal Survey of relevance to the Intelligence Community (IC) on emerging trends and future directions in the research area of social network thinking.1
Randolph H. Pherson, Pherson Associates, was asked to provide his perspective on the relevance of the presented research programs and ideas to the IC. As a consultant to the IC, he has had considerable experience working in teams and forming professional networks (or what he prefers to call “collaborative communities”) across agencies within the IC. He noted that all agencies within the IC have made considerable efforts to examine existing collaborative communities that appeared to work well and create new communities incorporating their features.
Pherson pointed out that some of the most effective communities were found to be those that drew their members from those who had worked together before; they could form a cohesive group quickly. Having members with networking skills that translated to connections with people and
1 For an archive of the full workshop Webcast, including discussion sessions, see http://sites.nationalacademies.org/DBASSE/BBCSS/DBASSE_181267 [December 2017].
resources outside the community was also helpful. He said he had found that communities that struggled to work together efficiently lacked mutual trust and had miscommunication problems, often as a result of being too large or too entangled in bureaucratic matters.
Pherson listed six imperatives he regards as necessary for professional networks to ensure effective collaboration and information sharing: (1) mutual trust, (2) mutual benefit, (3) incentives, (4) mission criticality, (5) access and agility, and (6) a common lexicon. He elaborated that for these communities, mutual trust refers to a willingness to demonstrate one’s own vulnerability; mutual benefit refers to everyone getting something productive out of the effort; incentives refer to support from management and leadership to engage in collaborative behaviors; mission criticality refers to recognition that collaboration should revolve around what team members must do every day and not be considered an add-on task; access and agility refer to permission and flexibility to participate; and a common lexicon refers to having common understandings of language, definitions, and rules of operation.
Pherson identified several areas in which further research on networks and network thinking could help inform the IC’s collaborative communities and their analysis of security issues. He suggested that more research could be conducted on the issue of trust within teams—how to build it and use it to bolster collaboration, as well as integrate new people into an existing team. He also highlighted the potential for research to address the challenges of information processing by developing tools to manage and help critically analyze the large amounts of information available, including the identification of any misinformation and information gaps.
Pherson expressed his view that the workshop presentations had focused attention on analytic areas worth the IC’s consideration, given that the future presents a different set of security problems than have been considered in the past. For example, he pointed to the applicability of the science of urban networks to models of political instability, as well as to humanitarian crises (e.g., current migration issues in Europe). He also suggested that the neuroscience work presented at the workshop would help clarify how like-minded networks develop, and could also enhance understanding of brain functions and subsequent behaviors in different contexts, such as when network interactions and communications take place in person compared with virtually. He speculated further that the concept of social identities presented by Kenneth Joseph could be applicable to deradicalization theories.
Kathleen Carley, Carnegie Mellon University, asked all the presenters to identify future research investments most likely to advance the science of social networks and social network thinking over the next 10 years and ensure its relevance and usefulness to the IC.
Regina Joseph, New York University, reinforced the importance of research on building trust within teams. She also suggested the need for research on identifying people with the right analytic skills and diversity of expertise to contribute to analytic teams within the IC in the evolving information environment.
Guido Cervone, Pennsylvania State University, asserted that research on human–machine interactions will be critical. He argued that since both the capability of the analyst and the output of automated decision-making algorithms are limited, efforts should focus on ways to combine the intuition of experienced analysts with machine outputs. He suggested that systems and professional networks be developed to take advantage of knowledge in different forms. This would require integrating the experience and knowledge of people possessing structured information from official data collections with the output of information generated by automated data mining systems.
Cervone commented on emerging technologies and their potential impact on research. He highlighted the increasing use of large combined datasets that include different types of data from such sources as remote sensing, numerical models, and social media. He also pointed out that although Cloud storage is nearly unlimited, data are accumulating at a much faster pace than the rate at which the data can be transferred to the Cloud and analyzed. The current trend in research, he said, is edge computing, whereby some of the research computation is carried out where the data initially reside. He mentioned two other technological advances: high-performance computing (e.g., exascale computing) and the ability to collect data in real time with one’s smartphone.
Jesse Hoey, University of Waterloo, pointed out that work in artificial intelligence is heavily data-driven. He argued for greater emphasis on building social or cultural models from such disciplines as sociology, psychology, and cognitive science to inform artificial intelligence systems. Kenneth Joseph, Northeastern University, expressed concern that some machine-learning approaches are starting to emulate human cognitive biases. He suggested that research explore how cognitive biases come to be embedded in data, how algorithms learn them, how this affects outputs from automated systems, and whether embedded biases in data can be leveraged by adversaries.
Markus Mobius, Microsoft Research, pointed out that little is known about individuals’ processing of communication and aggregation of information to take particular positions. He noted that while research can
measure outcomes and behavior, little is known about the internal processes involved. He drew attention to the longer life cycle of information and misinformation today relative to the past (i.e., people used to consume information in a newspaper in a day, whereas information now lives on for much longer in digital media). He also noted that in the current social media environment, information can appear as though it comes from multiple sources when it may in fact be from a single source.
Carolyn Parkinson, University of California, Los Angeles, said there has been some research on the structure of networks and the identification of network opportunities for communication. She suggested that future research investigate how different network structures impact information sharing and how a particular structure may impact the weight assigned to information or network members’ receptiveness to misinformation. She also suggested that research could be designed to examine people’s identification of a source of information, as well as their knowledge of that source and its relation to other sources, and how that source identification sways their positions or behaviors. In addition, she said, research could investigate how different people in structurally different network positions use their brains differently. Findings from this line of inquiry, she noted, could have consequences for identifying those who are more effectively targeted or influenced in certain situations. However, she cautioned, to understand individual differences using neuroscience techniques, a larger sample size is needed than that typically used for cognitive neuroscience experiments. She emphasized the promise of funding efforts that would bring together complementary expertise and create research teams able to integrate knowledge from psychology, cognitive neuroscience, sociology, social network thinking, and statistics.
Both Emily Falk, University of Pennsylvania, and Zachary Neal, Michigan State University, agreed that interdisciplinary teams will be important in the future. In response to a question on whether brain activity for decision making works the same way when messages are ill-intended (i.e., designed to get people to do something wrong instead of encouraging positive behaviors), Falk acknowledged that her work has focused on health-related behaviors and that neuroscience research on decision making in adults has generally focused on prohealth or prosocial choices. However, she noted, a large literature exists on risky decision making, particularly in teens and clinical populations. She suggested that research combining neuroscience and social network thinking could be expanded to consider brain functioning in decision making in response to ill-intended messaging. Neal, drawing on his experience in the field of urban networks, said he believes the research is tackling the right questions (e.g., Where do people, ideas, and money reside, and how are they connected? How and where do they move?). New questions are not needed, he asserted, but merging talent
across disciplines would enable the field to do a better job of answering existing questions. He highlighted as a challenge that each discipline brings its own specialization and language, such that researchers from different disciplines may be talking about the same thing without that being clear (e.g., the social sciences use the terms “nodes” and “edges,” while mathematics refers to “vertices”).
Hsinchun Chen, University of Arizona, suggested a different approach to the importance of interdisciplinary work—what he called a “three-leg approach to human data fusion.” In developing data mining tools for practical applications, he found that this work requires knowledge from the social and behavioral sciences to provide theory and explanations for social behaviors; analytic techniques from the data sciences to enable the aggregation of information from noisy data; and input from practitioners to provide insight on application and practical challenges. He emphasized that these three elements all need to be present to operationalize applications that involve data on human behavior and communications. He added that he sees more collaboration today than when he started in this area of research 20 years ago.
Benjamin Golub, Harvard University, called attention to research laboratories within commercial companies such as Microsoft. These labs, he said, follow a common approach of letting bright scholars work together in pursuing academic studies of interest to them within the companies’ domains of work. According to Golub, the studies are often innovative, and eventually companies can leverage some of the findings they yield.
A number of workshop participants pointed out that a large portion of social media data is controlled by private companies. Golub argued that a good, open data resource is needed for academics. Alexander Volfovsky, Duke University, agreed and asserted that no useful and easily accessible data are available with which to validate models. Chen noted that some longitudinal social media datasets are becoming available to social scientists and data scientists, although the data often lack identifying labels. Noshir Contractor, Northwestern University, added that some data that have been acquired from companies were not collected for research purposes but often come from server logs and were collected to help software engineers debug their platforms.
Contractor asked whether it would be possible to reengineer these platforms to collect data that would be more useful to researchers, and whether an agreement could be formulated that would allow the research community more regular access to social media data, sufficiently anonymized but still useful and labeled in some way. Volfovsky argued that it
may be more efficient to use by-product data than to create completely new pathways for data creation. He suggested gaining a better understanding of how and why the data are generated and embedding that knowledge in research models. On the other hand, he pointed to current experimentation on platforms that may collect the type of data useful for answering research questions. He also suggested that reaching agreement with companies for regular use of their data is unlikely because of the risk of research findings being misinterpreted and damaging to companies.
Golub asserted that any new set of data should be a “living” resource and not just a better dataset. He elaborated that regardless of how well a data collection is designed, researchers will continue to investigate different questions that will require different types of data in the future.
Marcus Mobius, Microsoft, noted that much of the data collected by companies has a short life. He highlighted data retention rules and legal requirements, especially in the European Union, that require companies to dispose of personal data after a certain period of time. He proposed developing anonymized, summary statistics to create a historical record of social activity without violating data retention rules. He acknowledged that some form of a consensus process would be necessary among the various research communities to determine how these summary statistics would be created and what information they would include. He pointed to Google Trends as an example of a successful statistical summary, one that has been used in many research papers. He proposed that an investment be made in developing more summary statistics for research instead of trying to determine how to transfer large amounts of data.
Scott Feld, Purdue University, offered two purposes for the research discussed at the workshop. The first, he said, is to understand networks in order to improve networks and build networks that effectively make progress toward goals. He offered the example of developing better teams in the IC to share and use intelligence information more effectively. A second purpose, he continued, is to understand how networks operate (e.g., how relationships develop and information spreads) in order to predict, control, or possibly disrupt behaviors.
Feld was asked to reflect on the ethics of conducting research on social networks. He defined ethics as “doing [one’s] work in an appropriate and socially acceptable way.” He explained that ethical concerns fall into three categories: (1) human subjects, (2) institution of science, and (3) responsibilities to the broader community and society. He stated that it is appropriate to worry about and protect from harm both people who
are direct participants in research studies (through informed consent and human subject approvals) and those whose personal information becomes part of studies, whether with their permission or covertly. He noted that researchers also have responsibility to their institutions and the institution of science, adding that they are expected to provide useful information, new knowledge, and generalizable results in exchange for infrastructure and funding.
Feld continued by asserting that research intended to inform the work of the IC or government must be undertaken with a well-defined understanding of that work. What this means, he said, is that data should be collected and theories considered that are relevant to the specific context at hand. He encouraged researchers to be humble and honest about what is known, what is not known, and any limitations of their research findings. It is important, he said, to think about what could be wrong with findings and why.
Feld raised a final ethical concern related to the effects on the community or society. He noted that research is often pursued for the public good, in an attempt to find optimal solutions to societal problems. For contexts involving an adversary, however, he raised the dilemma of the potential use of expertise to develop knowledge that may “cause harm in the interest of causing good.” He offered no easy solution to this dilemma, but emphasized the importance of considering the potential consequences of developing capabilities, knowledge, and sources of information that may impact society in negative ways.
Contractor and Carley closed the workshop by presenting their takeaways from the workshop presentations and discussions. Contractor focused on the tension between data-driven and theory-driven approaches to modeling. He underscored the importance of work on causal inference, which he characterized as an important methodological issue for moving the field of network science beyond describing networks to evaluating their processes and outcomes. He added that the type of research envisioned at the workshop will require interdisciplinary teams whose members have enough knowledge of each other’s areas of expertise to engage collectively in solving some of the research challenges highlighted at the workshop.
Carley commented that network science, much like statistics, is being projected to become part of every discipline, and the workshop had demonstrated the truth of this projection. She observed that the workshop had included much discussion of network metrics and the large amount of data available, but little discussion on interpreting and manipulating networks. She emphasized that although the amount of data has grown, the data are
still biased and incomplete, and she asserted that new tools for analysis are needed to address the uncertainties that result. She noted that the issue of cognitive bias had been mentioned; however, she called attention to the notion of stylized facts or known network biases. She suggested that scholars in the field of network science think about creating anthologies of robust findings in this area to help the IC and new researchers recognize what is implicit in the science of network thinking.