Automatic mapping and modeling of human networks
Introduction
A series of studies on office interactions discovered that 35–80% of work time is spent in spoken conversation, 14–93% of work time is spent in opportunistic communication, and 7–82% of work time is spent in meetings [1]. Senior managers represent the high end of these scales. Given the importance of such communications, it is notable that the majority of adults already carry a microphone and location sensor in the form of a mobile phone, and that these sensors are packaged with computational horsepower similar to that found in desktop computers. This emerging foundation of wearable sensing and processing power has allowed us to begin to automatically map and model how different groups within social or business institutions connect. We have been particularly concerned with our ability to automatically infer properties of human networks that affect propagation of information:
- •
Location context: work, home, etc.
- •
Social context: with friends, co-workers, boss, family, etc.
- •
Social interaction: are you displaying interest, boredom, friendliness, determination, etc.
By taking a statistical, machine-learning approach applied to the users’ behavior and physical situation, we have been able to show that it is possible to obtain solid, dynamic estimates of the users’ group membership and the character of their social relationships: e.g., who we work for versus those who work for us, or when we are interested versus when we are bored. By characterizing these patterns of behavior using statistical learning methods, we can then examine the users’ current behavior to classify relationships as workgroup, friend, interesting, and so forth.
The key to automatic inference of information network parameters is the recognition that humans are not general-purpose equipotent reasoning agents, but rather are creatures with a long evolutionary history that continues to shape our behavior and interactions with others. This shaping of behavior is particularly visible in social relationships and our attitudes toward them: we act differently when interacting with friends versus strangers, and when we are interested versus when we are bored.
Some of these categories can be inferred using standard methods such as surveys, however, these standard methods often suffer from subjectivity and memory effects, and their infrequency means that they are prone to becoming out-of-date. Even when information from standard methods is available, we would still like to use automatic methods to validate or even correct the standard information sources.
In this paper, we present statistical learning methods that use wearable sensor data to make reliable estimates about a user's interaction state (e.g., who was talking to whom, how long did the conversation last, etc.). We then use these results to characterize the connections that exist in groups of people.
Automatic mapping can be much cheaper and more reliable than human delivered questionnaires. For instance, in one of our studies we found that our automatic methods had an accuracy of 87.5% for detecting conversations with durations of 1 min or more. In contrast, a traditional survey of the same subjects produced only 54% agreement between subjects (where both subjects acknowledged having the conversation) and only 29% agreement in the number of conversations [2], [3].
Automatic discovery and characterization of face-to-face communication networks will also allow researchers to gather interaction data from larger groups of people. This can potentially remove two of the current limitations in the analysis of human networks: the number of people that can be surveyed, and the frequency with which they can be surveyed.
Automatic mapping of human networks will never be perfect, although it already seems superior to previous methods in some regards. We can also vary the confidence thresholds of the system, making the system more or less cautious about particular types of mistakes. In addition, the models provided by automatic mapping can suggest when traditional survey methods would be most useful, resulting in a semi-automatic capability that can have very high accuracy and a relatively low cost.
Section snippets
Socioscopes
Our approach to mapping and modeling human networks is to adopt the conceptual framework used in biological observation, such as is used to study apes in natural surroundings or in natural experiments such as twin studies, but replacing expensive and unreliable human observations with automated, computer-mediated observations. We imagined an advanced ‘socioscope’ that can accurately and continuously track the behavior of hundreds of humans at a time, recording even the finest scale behaviors
Reality mining
A critical requirement for automatic mapping and modeling of human networks is to learn and later categorize user behavior as quickly as possible. This is because the speed with which we can establish network parameters determines how accurately we can capture the dynamics of those networks.
Social signals
The importance of social displays has been highlighted by the research of Ambady and Rosenthal [13] and its practical ramifications explored in the popular book ‘Blink’ by Gladwell [14]. In brief, they have shown that people are able to ‘size up’ other people from a very short (e.g., 1 min) period of observation, even when linguistic information is excluded from observation, and that people use these ‘thin slice’ characterizations of others to quite accurately judge prospects for friendship,
Practical concerns
Continuous analysis of all interactions within an organization may seem unreasonable, and if misused, could be potentially dangerous. In an attempt to assuage some of these legitimate concerns, several methods of collecting this data will be discussed.
Conversation postings: In our experiments all the data were stored locally on the individual's machine. At the end of each day users could potentially review a summary of the number of conversations, the individuals involved, the character of the
Conclusions
These data make the point that human behavior is much more predictable than is generally thought, and is especially predictable from the behavior of others. This suggests that humans are best thought of social intelligences rather than independent actors, with individuals best likened to a musician in a jazz quartet. We can predict the behavior of these individuals from that of their associates because they are so attentive and automatically reactive to the surrounding group that they almost
Acknowledgments
Thanks to all of my collaborators and students for the hard work in forging these tools and this body of data. For future collaborators, you will find Matlab code, data, additional information and further publications available at http://hd.media.mit.edu. Portions of this paper have appeared in the Fifth International Conference on Development and Learning, Bloomington IL, May 31–June 2, 2006.
References (24)
Architecture and Communication Among Product Development Engineers
(1997)- S. Basu, T. Choudhury, B. Clarkson, A. Pentland, Towards measuring human interactions in conversational settings, in:...
- T. Choudhury, A. Pentland, The sociometer: measuring human networks, in: 23rd International Social Network Conference,...
- A. Pentland, Socially Aware Computation and Communication, IEEE Computer, March 2005, pp. 63–70. See TR 590...
- N. Eagle, A. Pentland, Social Network Computing, Ubicomp 2003, Lecture Notes in Computer Science, vol. 2864, Springer,...
- J. Gips, A. Pentland, Mapping Human Networks, PerCom ’06, Pisa, Italy, March 2006, see TR 586...
- A. Pentland, Social dynamics: signals and behavior, in: International Conference on Developmental Learning, Salk...
- A. Madan, A. Pentland, Vibefones: Socially Aware Mobile Phones, ISWC06, Martingy Switzerland October 11–15, 2006, see...
- N. Eagle, A. Pentland, Eigenbehaviors: Identifying Structure in Routine, October 2005, see TR 601...
- A. Pentland, T. Choudhury, N. Eagle, S. Push, Human Dynamics: Computation for Organizations, Pattern Recognition, vol....
Cited by (82)
Learning occupants’ workplace interactions from wearable and stationary ambient sensing systems
2018, Applied EnergyCitation Excerpt :The performance accuracy in detecting conversations was 63.5% overall and 87.5% for conversations greater or equal to one minute. To account for the small sensor variability and analysis methods, Pentland [11] describes analyzing wearable sensor data with several statistical learning methods such as principal components analysis and clustering methods in order to make reliable estimates of users’ interactions. He presents a detailed description of eigenbehavior modeling for learning and classifying user behavior from proximity and location data, and influence modeling for predicting the behavior of a subject from another subject’s data.
An android architecture for bio-inspired honest signalling in Human-Humanoid Interaction
2018, Biologically Inspired Cognitive ArchitecturesCitation Excerpt :Such conditions meet those that enable humans and animals to sense what is salient and act accordingly in a shared environment. Besides the theory of honest signalling has been already extended to human interaction laying the basis of sociometrics (Pentland, 2007). The paper is organized as follows.
Structural Analysis of the Stochastic Influence Model for Identifiability and Reduced-Order Estimation
2024, IEEE Transactions on Systems, Man, and Cybernetics: SystemsLearning ‘in the hive’: social character and student wellbeing in the age of psychometric data
2023, Critical Studies in EducationCOVID-19 Social Distance Proximity Estimation Using Machine Learning Analyses of Smartphone Sensor Data
2022, IEEE Sensors Journal