Behavior Informatics A New Perspective

Effective recommendation is becoming an in- Behavior Computing

creasingly important online and social behavior. Guandong Xu and Zhiang Wu share their view on Longbing Cao, University of Technology, Sydney

a topical issue—to involve, consolidate, and eval- Thorsten Joachims, Cornell University uate group preference for more targeted group- based recommendation.

Behavior is an increasingly important concept in Web search requests are more personalized and the scientifi c, societal, economic, cultural, political,

a context-aware understanding of information and military, living, and virtual world. In the dictionary,

62 1541-1672/14/$31.00 © 2014 IEEE

IEEE INTELLIGENT SYSTEMS

“behavior” refers to a manner of behav- price, and volume on a target secu- • Status (u)—where a behavior is ing or acting, and the action or reaction rity. The actions, response, or presen-

currently located; for instance, sta- of any material under given circum- tation, and the effect associated with

tus may refer to passive (not trig- stances. In Wikipedia, “behavior” refers the corresponding properties forms a

gered), active (triggered, but not to the actions and mannerisms made by concrete and rich object—the behav-

finished yet), or done (finished); in organisms, systems, or artificial entities ior of an individual or a group.

some other cases, status may in- in conjunction with its environment,

Behavior as a computational concept 1 clude valid or invalid. which includes the other systems or or- captures major aspects, including the • Associate (m)—other behavior in- ganisms around, as well as the physi- following: demographics of behavioral

stances or sequences of actions that cal environment. It’s the response of the subjects and objects; social relationships

are associated with the target one; be- system or organism to various stimuli or norms governing the interactions be-

havior associates may exist because a or inputs—whether internal or external,

behavior has an impact on another conscious or subconscious, overt or co- group; behavior sequences or networks

tween behaviors of an individual or a

or behaviors are related through in- vert, and voluntary or involuntary.

teraction and business processes to Thus, behavior is ubiquitous and fect generated by the behaviors on sub-

and their dynamics; and impact or ef-

form a behavior network. very social. In addition to the common jects and/or objects. Accordingly, an terms, such as consumer behaviors, hu-

Accordingly, a behavior instance man behaviors, animal behaviors, and ior (g) may carry (but is not limited to) (g) of an individual or group entity organizational behaviors, behaviors the following attributes and properties:

abstract and generic concept of behav-

can be represented in terms of a be- appear everywhere at any time. Behav-

havior vector → g as follows: iors in the physical world are explicit, • Subject (s)—the entity (or entities) and have been studied from many dif-

that issues the activity or activity → g = {s, o, e, g, b, a, l, f, c, t, w, u, m}. ferent aspects. With the rapid devel-

(1) opment and deep engagement of so- • Object (o)—the entity (or entities) cial and digitalized life with advanced

sequence.

Further, the behaviors of an individ- computing technology, in particular, • Context (e)—the environment in ual or group form a behavior sequence social networks, social media, online

on which a behavior is imposed.

G that can be represented in terms of a games, mobile applications, virtual re-

which a behavior is operated, in-

cluding the pre-condition and post- vector sequence →Γ, in which behav- ality, multimedia information process-

iors are connected in terms of social ing, visualization, machine learning, • Goal (g)—the objectives that the relationships R, and pattern recognition, more behav-

condition of a behavior.

behavior subject would like to ac-

iors in the virtual and social world are

→Γ = R{γ → 1, γ → 2, ..., γ → n}. (2) emerging. In addition, behaviors in tra- • Belief (b)—belief represents the in- ditional spheres are becoming increas-

complish or bring about.

With the vector-based behavior se- ingly complex with the involvement in

formational state and knowledge of

the behavior subject about the world. quences, further analysis on such vec- and marriage of the virtual and social • Action (a)—what the behavior sub- tors can identify vector-oriented behav- world. Socio-behaviors dominate these

ior patterns. Compared to traditional areas. Behaviors in more classic ar- • Plan (l)—the sequences of actions that sequential pattern mining, such vector- eas—including business, living spaces,

ject has chosen to do or operate.

a behavior subject can perform to oriented behavior pattern analysis is economics and politics—are also be-

much more comprehensive. coming more and more social.

achieve one or more of its intentions.

• Impact (f)—the results led by the

In different applications and sce- execution of a behavior on its ob- Computing Behaviors narios, behaviors present respective

Existing management information sys- social and non-social characteris- • Constraint (c)—what conditions tems and enterprise applications don’t tics, relationships, structures, and ef-

ject or context.

are imposed on the behavior; con- support the storage of behaviors very fects. For instance, in stock markets,

straints are instantiated into spe- well. The entity of physical or so-

cial behavior is usually decomposed ers and is embodied through trading • Time (t)—when the behavior occurs. to multiple transactions without pro- actions and action properties, such as • Place (w)—where the behavior tecting the semantics and complete placing a buy quote at a certain time,

a trader’s behavior influences oth-

cific factors in a domain.

happens.

behavior journey. Behavior as a very

JULY/AUGUST 2014 www.computer.org/intelligent 63

Behavior-oriented decision making and governance

Behavior presentation and management also provides a unified mechanism for describing and presenting behav-

Behavior reasoning

Measurement & evaluation

ioral elements, behavioral impact, and patterns.

• Analyzing behavioral impact. In ana- model checking

lyzing behavioral data, a person might representation & reasoning

pattern analysis intent analysis impact analysis

Behavior

Behavior

be particularly interested in those be- representation

Behavior

Behavioral data

Source data

learning & mining

havioral instances that are associated with a high impact on business pro-

Behavior-relevant applications & domains cesses and/or outcomes. Behavioral impact analysis 4,5 features the model-

Figure 1. Behavior computing research map. Behavior informatics consists of two

ing of behavioral impact.

major research directions: behavior representation and reasoning to formalize

• Discovering behavioral patterns.

behaviors, and behavior learning and mining to analyze behaviors.

There are in general two methods of behavioral pattern analysis. One is to

soft buzzword is used widely without

behavioral impacts, 4 and forming and

discover behavioral patterns without

a clear definition and systematic repre- decomposing behavior-oriented groups consideration of the behavioral im- sentation. Such “implicit” behavior in and collective intelligence for the emer-

pact, the other is to analyze the rela- transactional data isn’t consistent with gence of deep behavioral intelligence in

tionships between behavior sequences the “explicit” semantic existence in conjunction with their environments.

and particular types of impact. business. Hence, it’s necessary and cru- Behavior computing contributes to the • Emerging behavioral intelligence. cial to develop computing techniques in-depth understanding, discovery, ap-

To understand behavioral impact for explicit and in-depth quantification plications, services, and management of

and patterns, it’s important to scru- and informatics of behaviors.

tinize behavioral occurrences, evo- With the concept of behavior and

behavior intelligence.

lution, and life cycles, as well as the introduction of an abstract behav- addresses the following key aspects

In more detail, behavior computing

the impact of particular behavioral ior model, the representation, model- (as Figure 1 shows).

rules and patterns on behavioral ing, data analysis and mining, learning,

evolution and intelligence emer- and decision making of behaviors is be- • Extracting behavioral data. In pre-

gence. An important task in be- coming doable and increasingly useful,

havioral modeling is to define and essential, yet challenging in ubiquitous

paring behavioral data, behavioral

model behavioral rules, protocols, behavioral applications and problem-

elements hidden or dispersed in

and relationships, and their impact solving. They form a new computing

transactional data must be extracted

on behavioral evolution and intelli- opportunity, necessity, and technology

and connected, and further converted

gence emergence. innovation, which we refer to as be-

and mapped into a behavior-oriented

feature space, called a behavioral • Understanding behavioral network- havior computing or behavior infor-

ing. Multiple sources of behavior matics 2,3 for the explicit and in-depth

feature space. In the behavioral fea-

may form into a certain behavioral understanding and analysis of genuine

ture space, behavioral elements are

network. Particular human behav- behavior-oriented actions, operations,

presented into behavioral item sets.

Hence, it’s necessary to map and ior is normally embedded into such a and events associated with many chal-

network to fulfill its roles and effects lenging business problems.

convert transactional data to behav-

in a particular situation. Behavioral Behavior computing (or behavior in- • Representing and modeling behav-

ioral data.

network analysis aims to understand formatics) consists of methodologies,

the intrinsic mechanisms inside a net- techniques, and practical tools for

ior. This involves developing behav-

work—for instance, behavioral rules, representing, modeling, analyzing, learn-

ior-oriented specifications for de-

interaction protocols, convergence ing, discovering, and utilizing human, or-

scribing behavioral elements and the

and divergence of associated behav- ganismal, organizational, societal, arti-

relationships among the elements.

ioral item sets, as well as their effects ficial, and virtual behaviors, behavioral

The specifications reshape the be-

such as network topological struc- interactions and relationships, behavioral

havioral elements to suit the presen-

tures, linkage relationships, and im- networks, behavioral patterns, and

tation and construction of behav-

ioral sequences. Behavioral modeling

pact dynamics. 64 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS

Fundamental technologies

• Simulating behaviors. To understand

all of the aforementioned mecha-

nisms that may exist in behavioral Supporting data, simulation can play an impor-

techniques and tools tant role to observing the dynamics,

analysis

Formal methods

Social network

the impact of rules/protocols/pat-

System simulations

Organizational theor

Knowledge discover

terns, behavioral intelligence emer- gence, and the formation and dynam-

... ics of a social behavioral network.

• Presenting behaviors. From analyti- Logic

Theoretical foundation

Visualization cal and business intelligence perspec-

Multiagent systems tives, behavioral presentation aims to Cause-effect analysis explore the presentation means and ...

Sequence analysis

tools that can effectively describe the motivation and interest of stakehold-

Field structure: ers on the particular behavioral data.

Behavior informatics In addition to the traditional presen-

Mathematics

Systems sciences

Cognitive sciences

Social sciences

Information sciences

Intelligence sciences

tation of patterns (such as associa- tions), visual behavioral presentation is a major research topic. It’s of high

Figure 2. Behavior computing field structure. Behavior informatics, as a research

interest to analyze behavioral pat-

field, delves into three directions: theoretical foundations, fundamental

terns in a visual manner. technologies, and supporting techniques and tools. These tasks form a clear field struc- gain that can’t be achieved solely by • effectively capturing and quantify-

ture and research map of behavior com- usually recorded transactional data. The ing the relationships between be- puting. A generic process of computing deep values and prospects from comput-

haviors, behavior evolution, and behaviors will complement classic ap- ing behaviors may be through

their impacts, as well as measuring proaches toward a more comprehensive

the impact and performance of be- and in-depth behavior understanding • fully disclosing and utilizing the

haviors and behavior dynamics on and problem solving. Given a business

business objectives; application, it first converts entity rela-

behavior semantics that are usu-

ally destroyed in recorded transac- • deeply understanding the belief, de- tionship-oriented transactional data to

sire, and intent behind behaviors behavior feature-oriented data through

tions and overlooked in behavior

conducted and the impact caused; behavior modeling. Behavior patterns, • fully and deeply exploring the be-

analysis;

and

exceptions, dynamics, and impacts havior sequences and behavior ma- • actively detecting, early predicting, are then analyzed through developing

trix of an actor or a group along a

and intervening in unexpected be-

haviors of individuals, groups, or and learning methods. The outcomes

corresponding behavior-based analytic certain time period, in which be-

cohorts so as to convert them to the are then presented as behavior patterns, • deeply engaging in and learning

havior properties are involved;

expected directions and impact. rules, or visual diagrams, and/or trans-

about the explicit and (especially

formed into decision-support business

To access these prospects, cross-disci- rules to disclose the interior driving

hidden) social relationships gov-

erning behavior formation, struc- plinary efforts are needed. In addition forces and causes of business problems

turing, networking, evolution, and to informatics and analytics, theories, and impact.

methodologies, and tools available in • deeply discovering behavior pat- statistics, mathematics, econometrics, Prospects and Opportunities

emergence of behavior intelligence;

terns, exceptions, relational pat- marketing, psychologies, social science, Behavior is becoming an increasingly

terns, and changes of individuals, behavior science, behavior finance, and important asset to be deeply analyzed

groups, or the global population so on are necessary. This requires col- and understood to disclose its explicit

against behavior formation, evolu- laboration between disciplines and and implicit business value and semantic

tion, and revolution;

cross-domain experts (see Figure 2).

JULY/AUGUST 2014 www.computer.org/intelligent 65

T here exist unlimited opportunities Coupled Behavior

real-world applications, group behav-

in deep behavior computing in terms Representation,

ior interactions (that is, coupled behav-

of complementing the existing be- Modeling, Analysis, and

iors) are widely seen in natural, social,

havior analysis, data analysis, event Reasoning

and artificial behavior-related prob- detection, behavior economics, and

lems. Complex behavior and social ap- cognitive study towards data-driven, Can Wang, Commonwealth Scientific

plications often exhibit strong explicit semantics-oriented and process-based and Industrial Research Organization

or implicit coupling relationships both quantification and formalization of (CSIRO), Australia

between their entities and properties. what exactly takes place in the real Longbing Cao, University of

Moreover, it’s also quite difficult to world. Many application areas 2,3 from

model, analyze, and check behaviors traditional to emergent issues can ben- Eric Gaussier, University of

Technology, Sydney

coupled with one another due to the efit from it; for instance, exploring the Joseph Fourier

complexity from data, domain, con-

patterns, anomalies, sequencing of, Jinjiu Li, Yuming Ou, and Dan Luo,

text, and impact perspectives. and intent driving customer behav- University of Technology, Sydney

Due to the emerging popularity and iors in retail and online shopping busi-

importance of coupled behaviors, the nesses, Web usage, and interactions in Behavior refers to the action, reaction, representation, modeling, analysis, the Internet, trading behaviors in capi- or property of an entity, human or oth- mining and learning, and determina- tal markets, and exceptional activities erwise, to situations or stimuli in its tion of coupled behaviors are becoming captured on surveillance systems.

environment. 1 The in-depth analysis increasingly essential yet challenging of behavior has been increasingly rec- in ubiquitous behavioral applications References

ognized as a crucial means for under- and problem-solving techniques. They 1. L. Cao, “In-Depth Behavior Understand-

standing and disclosing interior driv- inevitably and undoubtedly consti- ing and Use: The Behavior Informatics

ing forces and intrinsic cause-effects tute new computing opportunities and Approach,” Information Science, vol.

on business and social applications, technological innovations, and thus 180, no. 17, 2010, pp. 3067–3085.

including Web community analysis, we refer to them as coupled behavior 2. L. Cao et al., eds., Behavior and Social

counter-terrorism, fraud detection, informatics, which is an important Computing, LNCS 8178, Springer, 2013.

and customer relationship manage- branch of behavior computing and an- 3. L. Cao and P.S. Yu, eds., Behavior

ment. With the deepening and widen- alytics. 4 Coupled behavior informatics Computing: Modeling, Analysis, Min-

ing of social/business intelligences and consists of methodologies, techniques, ing and Decision, Springer, 2012.

their networking, the concept of be- and practical tools for exploring hu- 4. L. Cao, Y. Zhao, and C. Zhang, “Mining

havior is in great demand to be con- man, organizational, artificial/virtual, Impact-Targeted Activity Patterns in

solidated and formalized to deeply qualitative, and quantitative behaviors, Imbalanced Data, IEEE Trans. Knowl-

scrutinize the native behavior inten- their interactions and relationships, the edge and Data Eng., vol. 20, no. 8, 2008,

tion, lifecycle, and impact on complex formation and decomposition of be- pp. 1053–1066.

havior-oriented groups, and collective 5. L. Cao, Y. Ou, and P.S. Yu, “Coupled

problems and business issues.

Although there’s an emerging focus intelligence.

Behavior Analysis with Applications,”

Here, we present the limitations IEEE Trans. Knowledge and Data Eng.,

on deep behavior studies, such as so-

cial network analysis, 2 periodic behav- of current research, and explore the vol. 24, no. 8, 2012, pp. 1378–1392.

ior analysis 3 and behavior informatics needs, opportunities, challenges, pros- approach, 1 previous research work has

pects, and trends of coupled behavior Longbing Cao is the director of the Ad-

informatics in terms of coupled be- vanced Analytics Institute and a professor

mainly focused on individual behaviors

without considering the interactions of havior representation and modeling as in the Faculty of Engineering and IT at the

them. However, with increasing net- well as analysis and reasoning. University of Technology, Sydney. Contact

work and community-based events

him at longbing.cao@uts.edu.au.

as well as their applications, such as Coupled Behavior group-based crime and social network Representation and

Thorsten Joachims is a professor in the De-

interactions, coupling relationships be- Modeling

partment of Computer Science and in the De- tween behaviors contribute to the in- Coupled behavior representation re- partment of Information Science at Cornell

trinsic causes and impacts of eventual fers to develop representation and University. Contact him at tj@cs.cornell.edu.

business and social problems. In the modeling mechanisms, languages, and 66 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS

JULY/AUGUST 2014 www.computer.org/intelligent 67

tools to capture behavior characteris- tics, intrinsic and contextual proper- ties of behaviors, behavior dynamics, and internal and external communica-

tions among behaviors. 5 Those tech-

niques and methods can also be used to understand interaction, causality, convergence, divergence, selection, decision, evolution, emergence, and intelligence of behavior entities, be- havior properties, behavior networks, and behavior impact. Both formal and visual specifications can be discussed to represent coupled behaviors and behavior interactions.

Limitations and Challenges

Existing behavior modeling approaches have too many styles and forms ac- cording to distinct situations. There’s very limited research on formalizing the concept of behavior and its ele- ments, which is too weak to reveal that behavior plays the key role of an inter- nal driving force for social and business activities. Additionally, it’s ineffective or even impossible to deeply tease out na- tive behavior intention and impact on complex issues and business problems. There are no formal behavior repre- sentation models stated from a gen- eral perspective and providing a com- prehensive understanding of behavior constitution.

In addition, state-of-the-art research work doesn’t explicitly model and an- alyze complex interactions of group behaviors directly. Complex cou- pling relationships between behaviors are often ignored or only weakly ad- dressed. Yet these behaviors are often observed to be correlated in terms of certain coupling relationships—for in- stance, serial or parallel, conjunction or disjunction. Such coupling relation- ships greatly challenge existing behav- ior representation methods, since they involve multiple behaviors from differ- ent actors, and add constraints on the interactions and behavior evolution,

which often aren’t obvious and exhibit large complexities. However, a deep exploration of interactive relation- ships is necessary for us to understand how behaviors are correlated and how those coupled behaviors drive and im- pact business and social problems.

Group behavior interactions, such as multi-robot teamwork and group communications in social networks, are widely seen in both natural, social, and artificial behavior-related applica- tions. Behavior interactions in a group are often associated with varying cou- pling relationships—for instance, con- junction or disjunction. Such coupling relationships challenge existing behav- ior representation methods, because they involve multiple behaviors from different actors, constraints on the in- teractions, and behavior evolution.

Research Objectives and Issues

Based on the aforementioned limita- tions and challenges, coupled behavior representation and modeling is to de- velop behavior-oriented specifications

and formalizations to describe coupled behaviors (that is, behaviors from ei- ther the same or different actors are often coupled with each other) and the relationships among them. It provides

a unified and formalized mechanism for describing, presenting, and aggre- gating behavior interactions, desired requirements or properties, behavior impact, and patterns.

Several classical theories and tech- niques are closely relevant to coupled behavior representation and mod- eling, such as ontology, knowledge representation, software engineering, cognition, agent, logic, and matrix computing. By taking great advantage of those underpinning mechanisms, the representation and modeling of coupled behaviors can be proposed, designed, and constructed in a solid and systematical way. During this process, a lot of research issues are

worth investigating, which include but are not limited to the following points:

• How can we define a unified con- cept of behavior? How can we de- fine a unified concept of coupled behaviors?

• What is a coupling relationship? How do we qualify and quantify the couplings or interactions among

a range of behaviors? • What kinds of basic coupling rela- tionships can we use to represent complex interactions among ho- mogeneous and/or heterogeneous behaviors?

• How do we model coupled be- haviors in both visual and formal manners? How do we establish a reversible and unique mapping or link between these two types of representation?

• What are respectively the syntac- tic and semantic interpretations of coupled behaviors? What is the re- lationship between them?

• How can we represent and abstract behavior interaction patterns?

Addressing these issues raises oppor- tunities for further analysis and rea- soning of coupled behaviors, which are widely seen in community and so- cial networks.

Coupled Behavior Analysis and Reasoning Coupled behavior analysis and rea- soning denote proposing effective methods, techniques, and tools for emergent areas and domains in an- alyzing and reasoning about cou-

pled behaviors and their properties. 1 Model checking technique is utilized to verify the coupled behavior model with desired requirements, and to further refine the model. Coupled similarities are also introduced to characterize the quantitative behavior

Individual

Coupled

analysis and reasoning, which are used Behavior

Behavior

reasoning

verification

to analyze, check, and verify complex behavioral elements, relationships, ag- gregations, properties, and constraints. It accordingly refines sensitive and problematic model proposals, and then

guarantees the robustness and stabil- representation

Behavior

Behavior

Behavior

integration

algebra

ity of coupled behavior representation schemes.

Likewise, strategies and theories in- cluding action reasoning and composi- tion, the belief-desire-intention model,

Behavior

Behavior

situation calculus, behavior composi-

learning

evaluation

tions, logic reasoning, and model check- ing have been proposed to analyze and reason about behaviors as well as their interactions. From the perspective of

Figure 3. Research issues on coupled behavior informatics. Coupled behavior

coupled behavior analysis and reason-

informatics aims to build systematic tools to address aspects and issues associated with individuals and groups with coupling relationships.

ing, there are many opportunities for us to widely explore. Many open issues are worth widely addressing and sys-

interactions in terms of coupling re- but on behavior exterior such as ser- tematically investigating. These inter- lationships between properties (such vice usage. The behavior implication esting research points include but are as attributes, features, and variables) in transactional data also determines not limited to the following: and/or entities (such as objects, re- that it fails to scrutinize behavioral in- cords, and observations). Algorithms tention and the impact on business ap- • The context of coupled behav- and case studies are discussed to an- pearance and applications.

iors is to be formalized to control alyze behaviors correlated with one

the whole process of coupled be- another based on mixed properties often overlooks the checking of be-

On the other hand, current research

havior analysis and reasoning ac- and complex coupled interactions. havior modeling, which weakens the

cording to different requirements. The analytical results will be used for soundness and robustness of models

More types of consolidated cou- detection, prediction, intervention, built for complex behavior applica-

pling are to be explored and stud- and grouping of coupled behaviors as tions. The quality of behavior inter-

ied, and the soft computing tech- well as their interactive relationships.

actions aren’t checked through verifi-

niques can be adopted to propose

cation techniques. Little related work

the fuzzy or rough coupled behav-

Limitations and Challenges

ior informatics. On one hand, traditional behavior verification of coupled behaviors, in- • Analytical problems such as the analysis is usually built on customer cluding elaborating and representing

is ready for the formalization and

convergence or divergence of cou- demographics and business usage-re- behavioral elements, specifying be-

pled behavior vectors are to lated transactions directly. It mainly havior-interactive relationships, and

be defined and intensively stud- relies on implicit behavior and explicit

ied. Limits of converged coupled business appearance from behavioral havior couplings. The engagement of

checking the modeling of multiple be-

behavior sequences are to be clari- and social sciences, often leading to in-

fied, which are essential to calculus effective and limited analysis in under- make the findings much more stable

verification in behavior analysis may

and can be used to define continu- standing business and social activities and robust for problem solving.

ity, derivatives and integrals. deeply and accurately. With behavior

• Some other research issues, includ-

implied in demographic and transac- Research Objectives and Issues

ing how to define the bases and di- tional data, it’s not possible to support With the formal representation of cou-

mension of such a coupled behavior in-depth analysis on behavior interior pled behaviors, the coupled behavior

space, how to do the space decom- surrounded by behavioral elements, analytics addresses the task of behavior

position, how to conduct the linear 68 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS

JULY/AUGUST 2014 www.computer.org/intelligent 69

and nonlinear transformations, are to be addressed and deeply explored.

• How can we use logic-related repre- sentations to reason about coupled behaviors and their interactions? How can we adopt verification tech- niques to check the validity of cou- pled behaviors via certain proper- ties? How can we extract rules and mine patterns from analyzing and reasoning about coupled behaviors?

As a result of exploring these research issues, we’re able to develop a deeper understanding of behaviors, especially coupled behaviors of interested groups.

B ased on these discussions, we sum- marize the associated research issues

for coupled behavior informatics (see Figure 3). Here, we mainly focus on coupled behavior representation, cou- pled behavior reasoning, and coupled behavior verification. In fact, these points are designed for qualitative coupled behaviors, which are qualified by actions. Alternatively, some coupled behaviors are quantified by properties, called quantitative coupled behaviors. Accordingly, coupled behavior learn- ing and coupled behavior evaluation are proposed for the quantitative cou- pled behaviors. Finally, a coupled be- havior algebra can be introduced to integrate the qualitative coupled be- haviors and quantitative coupled be- haviors, which forms a big picture for coupled behavior informatics.

References

1. L. Cao, “In-Depth Behavior Understand- ing and Use: The Behavior Informatics Approach,” Information Sciences, vol.

180, no. 17, 2010, pp. 3067–3085. 2. T. Hogg and G. Szabo, “Diversity of Online Community Activities,” Proc. 19th ACM Conf. Hypertext and

Hypermedia, 2008, pp. 227–228.

3. L. Cao, Y. Ou, and P. Yu, “Coupled Behavior Analysis with Applications,” IEEE Trans. Knowledge and Data Eng., vol. 24, no. 8, 2012, pp. 1378–1392.

4. L. Cao and S.Y. Philip, Behavior Com- puting: Modeling, Analysis, Mining and Decision, Springer, 2012.

5. C. Wang and L. Cao, “Modeling and Analysis of Social Activity Process,” Behavior Computing, 2012, pp. 21–35.

Can Wang is a postdoctoral fellow with Commonwealth Scientific and Industrial Research Organization (CSIRO), Australia. Contact her at can.wang@csiro.au.

Longbing Cao is the director of the Ad- vanced Analytics Institute and a professor in the Faculty of Engineering and IT at the University of Technology, Sydney. Contact him at longbing.cao@uts.edu.au.

Eric Gaussier is a full professor in computer science at the University of Joseph Fourier. Contact him at eric.gaussier@imag.fr.

Jinjiu Li is a lecturer in the Advanced Analytics Institute and at the University of Technology, Sydney. Contact him at jinjiu.li@uts.edu.au.

Yuming Ou is a lecturer in the Advanced An- alytics Institute and at the University of Tech- nology, Sydney. Contact him at yuming.ou@ uts.edu.au.

Dan Luo is a research fellow in the Ad- vanced Analytics Institute and at the Uni- versity of Technology, Sydney. Contact her at dan.luo@uts.edu.au.

Behavior Analysis in Social Media

Reza Zafarani and Huan Liu, Arizona State University

With the rise of social media, informa- tion sharing has been democratized. As

a result, users are given opportunities

to exhibit different behaviors such as sharing, posting, liking, commenting, and befriending conveniently and on

a daily basis. By analyzing behaviors observed on social media, we can cat- egorize these behaviors into individ- ual and collective behavior. Individ- ual behavior is exhibited by a single user, whereas collective behavior is observed when a group of users be- have together. For instance, users us- ing the same hashtag on Twitter or migrating to another social media site are examples of collective behavior. User activities on social media gener- ate behavioral data, which is massive, expansive, and indicative of user pref- erences, interests, opinions, and rela- tionships. This behavioral data pro- vides a new lens through which we can observe and analyze individual and collective behaviors of users.

The emergence of this new type of data presents behavior analysis on social media with new challenges. We detail first what individual and collective be- havior analysis is, and then outline novel challenges with future work.

Individual Behavior Analysis Individual behavior can be consid- ered one of the following:

• User-user behavior. This type of behavior is observed between two users. For example, befriending and following in social media are examples of such behavior.

• User-entity behavior. This type of behavior is exhibited with respect to entities on social media (for ex- ample, user-generated content). For instance, liking a post on Facebook or posting a tweet on Twitter are examples of user-entity behavior.

• User-community behavior. This is the type of behavior that users ex- hibit with respect to communities. Joining and leaving communities are examples of this type of behavior.

70 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS

Irrespective of the type of behavior observed, we can utilize a computa- tional methodology to analyze behav- ior and help find interesting patterns as follows: 1,2

To analyze individual behavior, we can collect traces of the behavior on so- cial media. For instance, for analyzing the user-befriending behavior, we can collect the list of individuals that users befriend over time. To understand the underlying reasons for befriending, we can follow a machine learning tradition and create data features that are likely to be related to the behavior. For the befriending behavior, we can consider the number of common friends as an important feature. After constructing the features, we can find the correla- tion between features and behavior us- ing a supervised learning framework. Since correlation doesn’t imply causa- tion, evaluation techniques are required to verify the validity of our findings. We

can use randomization tests 2 or causal-

ity testing techniques, such as Granger’s causality 2 for evaluation purposes.

Collective Behavior Analysis Collective behavior analysis can be per- formed by either analyzing the individ- uals that exhibit the collective behav- ior independently, or by analyzing the individuals that exhibit the collective behavior as one group. In the former, we’re aggregating (summing, averaging, taking majority, and so on) the result of individual behavior analysis, which can

be performed using the aforementioned methodology. In the latter, we consider the individuals exhibiting the collective behavior as one (large) group, and the behavior is analyzed for this group. As the focus is on group-level behavior, we can use methods that model group-level dynamics to analyze collective behavior. For instance, epidemic models from epi- demiology or techniques that analyze influence on implicit networks can be used to analyze collective behavior. 2

Case Studies Let’s discuss four behavior analysis case studies. Two are examples of individual behavior analysis (community mem- bership behavior and connecting users across sites) and the other two (movie revenue prediction using Twitter and user migration in social media) are in- stances of collective behavior analysis.

Community Membership Behavior

Users often join communities for dif- ferent reasons. To analyze this be- havior, Lars Backstrom and his col-

leagues 3 collected information about users joining communities over time and designed features that could have influenced users joining communi- ties. They determined how impor- tant these features are in predicting whether users join communities by us- ing a decision-tree learning algorithm. Their findings suggest that not only is it more likely for individuals to join communities when they have many friends inside the community, but it’s also important how these friends are connected inside these communities— for example, how dense their friend- ship network is.

Connecting Users across Sites

In previous work, 4 we connected us- ers across social media sites by identi- fying multiples sources of information that are generated by the same user. We noticed that the minimum information available on different social media sites is the username individuals select. By employing usernames alone, we iden- tified profiles that represent the same individuals across social media sites. We analyzed behaviors that individu- als exhibit when selecting usernames, such as selecting the same usernames, using the same language or vocabulary, and their typing patterns, among other behaviors. These behaviors were cap- tured using data features and helped successfully connect users across sites.

Movie Revenue Prediction Using Twitter

Sitaram Asur and Bernardo Hu- berman 5 attempted to predict the collective behavior of going to the movies by analyzing the traces it left in the microblogging site Twitter. They found that by em- ploying only eight features, movie revenue can be predicted with high accuracy. These features are the av- erage hourly number of tweets re- lated to the movie for each of the seven days prior to the movie open- ing (seven features) and the number of opening theaters for the movie (one feature).

User Migration in Social Media

Working with Shamanth Kumar, 6 we analyzed the collective behavior of users migrating across sites. We showed that using three general fea- tures that measure user’s activity, us- er’s network size, and user’s prestige, we can effectively model and predict populations that migrate across so- cial media sites.

Behavior Analysis Challenges User behavior analysis in social me- dia faces stern challenges. Here, we outline some immediate and demand- ing issues.

Data Sparsity

Not all behaviors are easily observ- able on social media. Consider ana- lyzing the money individuals spend on social media or their driving routes. These data aren’t as abun- dantly available on social media as they are in the physical world. In other words, while for specific pat- terns (such as befriending) massive sources of data are available on so- cial media, for other behaviors data are sparse. This imbalance in data availability limits the behaviors that

JULY/AUGUST 2014 www.computer.org/intelligent 71

can be analyzed using social media and at the same time provides op- portunities for identifying relevant information sources for behavior analysis.

Lack of Causality Information

Behaviors in social media are only observed by the traces they leave in social media. We rarely observe the driving factors that cause these behav- iors; nor can we interview individuals regarding their behaviors. Consider

a tweet of John that is retweeted on Twitter by Mary. Does that mean that the tweet is interesting to Mary? Does that mean that the tweet is worth spreading to others? Or, does that mean that John and Mary are friends and Mary retweets all tweets by John? These interesting questions are fre- quently encountered when trying to identify causes of behavior in social media.

Evaluation Dilemma

Even if a behavior is analyzed on so- cial media and related patterns are gleaned, it’s difficult to verify the va- lidity of these behavioral patterns. Evaluation becomes even more chal- lenging for industries in which impor- tant decisions are to be made based on observations of individual behav- ior. For instance, consider an online seller that observes an abnormal rise in their site visits and sales. The site can expand and invest in the infra- structure to handle such traffic. How- ever, this traffic could be due to a sudden demand; hence, temporary. A deeper understanding of its members is necessary to ensure the validity of these behavioral patterns.

T he journey of understanding hu- man behavior at scale has just begun.

Most current work in social media

considers analyzing behavior from

a data mining or machine learning perspective. Social sciences, includ- ing psychology, anthropology, and ethnography, have developed their theories and rigorous methods to un- derstand human behaviors in small groups and at a small scale. It’s im- perative to generalize these theories to larger populations observed in social media and to design new techniques tailored to behavioral analysis on so- cial media.

References

1. L. Cao, “In-Depth Behavior Un- derstanding and Use: The Behavior Informatics Approach,” Information Science, vol. 180, no. 17, 2010, pp. 3067–3085.

2. R. Zafarani, M.A. Abbasi, and H. Liu, Social Media Mining: An Introduction, Cambridge Univ. Press, 2014.

3. L. Backstrom et al., “Group Formation in Large Social Networks: Membership, Growth, and Evolution,” Proc. 12th ACM SIGKDD Int’l Conf. Knowledge Discov- ery and Data Mining, 2006, pp. 44–54.

4. R. Zafarani and H. Liu, “Connect- ing Users across Social Media Sites: A Behavioral-Modeling Approach,” Proc. 19th ACM SIGKDD Int’l Conf. Knowl- edge Discovery and Data Mining, 2013, pp. 41–49.

5. S. Asur and B.A. Huberman, “Predict- ing the Future with Social Media,” Proc. Web Intelligence and Intelligent Agent Technology, vol. 1, 2010, pp. 492–499.

6. S. Kumar, R. Zafarani, and H. Liu, “Un- derstanding User Migration Patterns in Social Media,” Proc. AAAI, 2011.

Reza Zafarani is a doctoral candidate in Computer Science and Engineering at Ari- zona State University. Contact him at reza@ asu.edu.

Huan Liu is a professor in Computer Science and Engineering at Arizona State University. Contact him at huanliu@asu.edu.

Group Recommendation and Behavior

Guandong Xu, University of Technology, Sydney Zhiang Wu, Nanjing University of Finance and Economics

With the popularity of the Internet and the use of database systems (or various information systems), the in- teraction between the system and its users, which conveys valued user be- havior information, has become a dominant observation in the current information era. This triggers a de- mand to develop effective and effi- cient techniques and systems to cap- ture and analyze the user behavior patterns, in turn facilitating other intelligent Internet services. Recom- mender systems (RSs) have emerged as a powerful tool for helping peo- ple find relevant contents from vast

amounts of information. 1 From the perspective of applications, RS is a typical extension of behavior analyt- ics in information provision. To date, numerous techniques have been de- veloped for both efficient and effec- tive recommendation, and they have been widely used in not only research but also real e-business applications, such as the ones used by Amazon and Netflix. The essence of RSs is how to extract, model, represent, and uti- lize the user behavior pattern induced from human-computer interaction activities effectively and efficiently.

Much of the current focus within RSs has been on making recommen- dations for individual users; in this sense, it’s also often called personal- ization or personalized recommenda- tion. However, in some real-life cases the items to be recommended are not for personal use but for a group of users. For instance, a DVD could

be watched by a group of friends or in a family. On the other hand, the be watched by a group of friends or in a family. On the other hand, the

Different from personalized recom- tings, 4 such as the analysis of demo- aggregated ratings for the group. This mendation, group recommendation graphic, behavioral, and transactional method was used in both the PolyLens aims to

data. For example, weblog analysis (or and Movie Lens systems. The last one customer purchase basket analysis), is somehow utilizes information about the

• reveal the hidden communities (or

a typical type of user behavior pattern preferences of individuals to arrive at a groups) from the whole population, analysis, where similar user interests model of the preferences of the group as • represent the group profile to re- or tastes can be captured, and user

a whole. Two typical systems using this flect the common user interest or profiles can be utilized. A further ap- method are Let’s Browse and Intrigue. preference, and

plication of such behavior analysis is

• eventually make recommendations the recommendation, which aims to Score Aggregation

by taking into account the global predict new products or services that The score aggregation approach com- relevance of items to the group’s users may be interested in based on putes each member’s individual rec- preference and the minimum dis- the patterns learned from behavior ommendations and merges them to agreement of items within the analysis. In this sense, recommender produce a single list for the group, group.

systems have a close connection to be- where the score of each item is ag- havior analysis in terms of behavior gregated from individual recommen-

For this reason, there is a grow- pattern analysis, behavior represen- dations. Two main score aggrega- ing interest in group recommenda- tation, and behavior use. Next, we’ll tion functions have been presented tion technologies that identify recom- discuss the technical aspects of group thus far: average and least misery. 2 mendations to satisfy the individual recommendations addressing these The average function aims to maxi- preferences of all users in a group as three issues.

mize the average of the group mem- much as possible. 2 Here, we present a

ber’s scores for each item, and it was review on the state of the art in this Group Recommendation

employed by some practical systems area. Because community detection is Techniques

such as Pocket and RestaurantFinder. an important topic in social network Almost all of the existing methods for Meanwhile, the least misery func- analysis, the focus of this article will group recommendation target aggre- tion targets maximizing the lowest

be concentrated on the representa- gating individual group members’ rel- score among all group members. The tion and utilization of group behavior evances to generate recommendations.

PolyLens adopted this strategy. patterns.

No matter which strategies are ad- part of group recommendation, can opted, the key point among these ap- Behavior Analysis and

The aggregation technique, the core

fall into two categories: preference ag- proaches is to aggregate the user behav- Recommendation

gregation and score aggregation. 2 ior representations via accumulating the From cognitive perspective, behav-

rating values or recommendation scores

ior represents the action or reaction Preference Aggregation

of items for the group. From this per- of an entity or human, to situations Preference aggregation is to combine spective, the essence of group recom- or stimuli in its environment, which each individual member’s prior rating mendation is the group behavior or can be pervasively seen anywhere at into a single overall rating for the group preference aggregation. In this context, any time. More specifically, it refers (group profile), while recommendations the group behavior or preference pro- to “those activities that present as are made to users. Commonly, there are file is analogous to the item set or rating

actions, operations or events as well three schemes to be used by the meth- scores, reflecting how the group users as activity sequences conducted by ods along this line. The first one is to will perceive or prefer the target items. entities within certain contexts and merge recommendations made for in-

However, the aforementioned two environments in either a virtual or dividuals (for example, one alternative mechanisms only consider the overall

72 www.computer.org/intelligent IEEE INTELLIGENT SYSTEMS

JULY/AUGUST 2014 www.computer.org/intelligent 73

relevance of items to group members. In real-world cases, an item might have different relevances to different group members and this disagree-

ment among members must be fully taken into account. Taking a closer look, the former concerns the overall consistency of relevances within the group, while the latter emphasizes the relevance variances of items among each member in the group.

Consensus of Group Preference