Analysis of a Social Activities Recommendation Model
Sasha Mile Rudan
Oslo University +47 41 753 707 sasharu@ifi.uio.no Sinisha Rudan Belgrade University +381 64 28 30 738 sinisarudan@gmail.com Aleksandar Cikaric Belgrade University +381 63 345 546 cikaric@gmail.com
ABSTRACT Online Social Networks have become a place where much more than social interactions and friendships are happening. Over the last few years more and more real business applications have migrated to the online communities, and in some sense power structures have migrated to the online virtual communities. However neither individuals are presented with the power they can harvest through the more efficient social network usage, nor they are faced with the network topology that can give them insights. In this paper we present our research on Social Activities Recommender (SAR) model, a proactive social-network model that suggest optimal activities based on Trans-Technical Patterns that are captured either through user interview and tunings done by users (iteratively through the usage period) or described by domain experts as a best practice and later even (manually) monitored by domain experts. We show that little as 10% of suggestions (in that sense sacrificing only 10% of a free own will) a person can increase efficiency for about 50% of standard and estimated maximum. The use of WhoAmI platform (users preferences modeling system) is essential for this research. WhoAmI shapes, or in other words, it builds the landscape of the social-network (a way to understand actors and their needs/interests) we are analyzing. It also shapes the data-flow happened between actors, and the last, but probable the most important, the data/knowledge visible to each actor. We present our research in the context of real-world example, Activity Buddies; an online social network system for supporting offline (real-world) human activities. INTRODUCTION Research conducted for this paper relates to a social network Activity Buddies (AB). Authors of this paper are working as researchers and partial architect managers of AB. AB is a service that provides online support of offline (real-world) human activities. Intention is to promote activities, to engage members to participate in more activities and to support serendipity in order to discover new activity partners. AB focuses on real-time, responsive, at-place ad hoc arrangements, supporting social-growth theory rather than activities organization and management. WhoAmI is a key component necessary for understanding users, users connections, and interests, and in particullary interactions, since interactions are mostly covered through trans-technical patterns, rather than WhoAmI profiles. AB system is an example of a Trans-Technical System (TTS) and therefore is modeled to support trans-technical- processes (an extension of social processes), which means that system will be developed in service-oriented style (SOA) with system components that simply provide features required for particular system domain. However, system usage, user interaction, and design are left out of the the features architecturing. This is an important production detail, especially if we think of small smart mobile devices which screen is small and user experience suffers much more in the case of components and details of no interest. The second important aspect of our research and change in SNA (Social Network Analysis) application is integrated in TTS systems and therefore is a part of the AB system as well. Opposite to many others cases of SNA application we have shifted SNA application from simple passive and observatory role toward more proactive and dominant role. Under this light, we introduce a trans-technical pattern (TTPt) as a fundamental component of proactive SNA. TTPts are created either by human domain experts or automatically by system, through observation and learning. The general model that will be researched, modeled and implemented is the following: 1) system actors (members) are describing their objective interests through the system, 2) domain experts are describing trans- technical patterns (TT-patterns) that are describing best-practice objective methods, 3) system is observing system actors behaviors and by using TT-patterns it can recognize critical difference in actual (behavioral) results and expected (objective) goals and 4) system proactively promotes corrections through suggestions to actors, finally 5) actors responds to suggestions and corrects behaviour and in that sense correlate more with initial goals. An important note here is that the SAR system will depend on underlying social-network topology and real-time values that describe the network. Values will describe a longitudinal social-network structure, where edges will be a superposition of micro-activities marshaling between actors. In the following chapters we will introduce Activity Buddies system and its components together with the more general TTS system, and finally explain concepts of proactive network components inside of them and benefits of using them. ACTIVITY BUDDIES Activity Buddies (AB) is a social network that aims to transparently support real-world activities and spontaneous serendipity of matching activities. It tries to mimic and understand social processes in the real world and to mimic the real-world as much as possible (Chan 2004). In order to achieve the abovementioned aims, AB design and functionality (from the control flow perspective rather than from the set of supported features) is not dictated by features that AB creators wanted to incorporate in it or by their understanding of the real-world functions, but on the contrary, it utilizes TTS systems paradigm and offers lego boxes that users use and organize in order to assemble their own needs and replica of the real-world projection they care of. That act of users deciding on a structure of social processes they are utilizing on the top of the AB system and way how and what system blocks they want to incorporate in their system usage practice is one of the essential ideas of TTSs and a strong guidance for a novel approach in developing AB system. To be possible to incorporate TTS paradigms, there are few important aspects of AB modeling that needs to be taken care of. System learns and adapts itself to user custom world representation, application usage approach, requirements, interaction and so on, instead of vice-versa. Our research is driven by requirement of designing AB as a set of components able to be represented as Trans- Technical Components (TTC). This means that relationships between components should not be functionally dependent, in other words. components should avoid even loose coupling, i.e. functional dependencies, and allow coupling either through data or functional flow only at Trans-Technical Processes (TTPs) level (Fig 1). It might seem that this restriction is too expensive or that it places too much anarchy in the system control flow, however that is not necessarily true. TTPs can enforce unique path or have more favorited paths, both supported by trans-technical patterns (TTPts). In that way, domain experts are able to encode necessary functional control flows in the system, but provide it at the level of user TTPs which makes it more easy to redesign and therefore support lean/agile development in very responsive way (Wagner 2000) and (Taveter, 2001). Even if that alone seems to be very valuable reason and investment for shifting toward TTSs, we believe the major benefit is in possibility of lifting all hardcoded wiring in the system to higher soft descriptive level. Besides of the possibility of easier system rewiring, this approach puts on the same level system experts, domain experts, and users, giving a chance for mashing up and mixing their preferences in the best possible way. On the top of that, we are able to develop special system components that will observe TTS and user behaviors, learn, monitor and react (all through TTPs) to support more efficient TTS usage outcomes.
Figure 1. A generic example of Trans-Technical process (TTP) We would also need to change graphical and interactive design practice in order to make it more supportive of TTS specific characteristics, and make dynamic control/data flows easier for adoption and make them more apparent to users with intention to use the visualized flows as a navigation interface through the complex system workflow. However, that is part of a later research. To make it possible that TTS supports free data and control flows, encoded through TTPs, we need to represent each data artefact / entity as a successor of Trans-Technical Entity (TTE). TTE is an object that can be operated through the application of TTP and it can be a part of TTPts. Finally, all TTEs need to be interconnected with all other entities (even if the strength of relationships in most cases equal to 0.0) in order to support continuous transitions and serendipity recommendations (Fig. 2).
Figure 2. Friendships and WhoAmI (Activity Buddy System) With all these aspects taken care during TTS system development (AB, in this particular case), we end up with a system that is open for TTP modeling and monitoring. We are able to introduce SNA, ABM, and several other relevant components which through the presented architecture have a full access to TTEs, system components (services), machine learned TTPts. This gives possibility to any system/domain expert to take active control over the system behavior and how system interacts with systems users. System interact with its users through proactive components, the after integration in the system can react based on acquired knowledge (own or from other components) and change trans-technical patterns (TTPs in bigger picture); effectively inhibiting or fostering particular behaviors (Liu 2008). WhoAmI SERVICE Over the whole decade of online social networks existence and semantic web, the semantic representation of user profiles still didnt get stronger momentum in any popular systems. In the similar way, social inter-communication and communication in general didnt get filtered or shaped by personal profiles/interests. WhoAmI service is a service addressing above mentioned issues, by providing a semantic interpretation of someones interests.
Figure 3. WhoAmI model representing 3 interests WhoAmI is a service helping individuals to explain their interests, build reputation and to match on the base of common interests. Each actors interest is represented as a WHO(AmI) object, which is defined by: WHAT - an interest user is relating himself to (e.g. HTML5, VHDL, volleyball, dance, writing, etc.) HOW says how he is related to it (e.g. like, practice, teach, etc.) WHERE defines privacy/visibility rules (defines relationships to one of user roles, profiles or apps) TTS and TTS-LA In this work we model TTS as a SOA (Service Oriented Architecture) systems. It helps us to rely on existing standards like SoaML (Service oriented architecture Modeling Language). Usages or references of existing standards will help us to model systems in a recognized way, to benefit from existing tools, and to have higher acceptance ratio both in academic and industry domain.
Figure 4. Trans-Technical System Layered Architecture (TTS-LA) Trans-Technical System - Layered Architecture (or shortly TTS-LA) is an architecture, which makes us possible to support SOA system model. It also helps us to provide integration of 3rd party components and wrap them in a unified model and unified/federated knowledge among components and users (Chiu 2006). However, the most interesting part of TTS-LA for this particular research is happening on one higher level, where we have a set of monitoring components that are observing users and system interactions and that are building knowledge on the top of observations, ending up with a set of Machine Learned TTPts. Machine Learned TTPts add up to the set of experts created TTPts and help/suggest/guide user actions to help users to increase their or system overall efficiency. TRANS-TECHNICAL PATTERNS Having set up AB and TTS system integrated with WhoAmI, we are able to present the more advanced components and benefits of introduced models and concepts. Describing Users Through Patterns Apart from TTPts manually described by experts (which describe a generalized user, a domain and system patterns) and machine learned patterns (in form of discovering new specific and possibly generalized patterns), important type of patterns are those that we are building through interviewing users or observing users behavior in predefined and well know scopes. Proactive SNA Majority of SNA applications come as an off-track side tool for network analysts to analyze and most often to describe and understand social behaviors. However, rarely are SNA tools integrated as a part of the system where they can be supported by user and system knowledge (Cramton, 2001) in order to predict and direct users and systems behaviour and destination state. This is the main purpose of integrating proactive SNA components in our TTS system (Fig. 5.)
Figure 5. Proactive component of TTPs and SNA+ABM The main purpose of proactive SNA (pSNA) is to use SNA to analyze particular users-system state and through the use of TTPts (inhibitors and promoters) to shift addressed (or the most critical) part of the users-system toward more optimal state. It can also protect regular users against Sybil profiles or even more subtle undesired scenarios (Levin 2004). Inactivity ring AB modeling provides members by benefits of exploiting structural holes and social capital. Extreme example, that we have identified in our research is the inactivity ring in which persons (person A), willing to be active, are preserved inactive by being surrounded with inactive friends. We focus on finding ways to break this ring. One of our approaches is by using SNA tools to identify active persons around this ring and to identify gatekeepers B that can integrate active persons C and members of their clique to activate inactive ring. Prior this integration, payoff (in the aspect of Game theory) for anyone in inactivity ring for being active was higher than for being inactive. After it, payoff gets higher and higher, in an iterative way, and motivation spreads, as more persons in the ring engage in active socialization (hanging out) (Sekiguchi & Nakamaru 2011).
Figure 6. Inactivity Ring Teams creation suggestion Our research shows that often inactivity preserves because people are not aware that there are people/friends around them willing to share the same (or a new) activity. Simple visualization of these persons with shared interests, performed by AB, can motivate them, but yet we go step furthers. We periodically and automatically offer them, joint activities. Even further, AB is using SNA tools to identify connected components in the system whose members are inactive and are geographically not far from each other. AB examines their interests and offers them activities (sport/outdoor) that are being identified as interesting for most of the component members. System offers them entering the group it has automatically created for them, and later suggests events creation and venues for those events. Additionally, AB identifies external motivators - persons out of the connected component, but being connected to one of its members and being active in the suggested activity, in order to additionally boost newly created group activity. In such created teams AB identifies possible leaders through analyzing members Betweenness and Closeness Centrality and through analyzing members actions, and then motivates them and suggest them leading roles and activities. DNA - Dynamic network analysis Besides the static approach, AB performs Dynamic network analysis in order to detect changes in members and group activities that may be manifestation of unhealthy patterns. System takes proactive role and rises warnings when it detects it. Even further, we are performing research in understanding of processes that lead to unhealthy states in online communities / social networks. Understanding of these processes and symptoms is at the stage of understanding human anatomy and illness processes prior of Leonardo da Vincis dissection of human corpses. Thus, our research is focused on dissection of extinct and dying out or ill networks in order to notice patterns and symptoms. AB further tracks its communities topology and dynamics, comparing it to gathered patterns from already examined ones to predict unwanted community states and processes. In the same time, it detects and gathers new patterns from the examined communities. ABM Modeling Another interesting component we can see at the Fig. 5. is an ABM (Agent Based Modeling) component. Activity Buddies research and development is interesting by our approach to ABM methodology, by not using it for an offline model simulation or prediction only, but by introduction of its using for real-time integration and prediction of close and distant future of users-system states. We examine modes of usage of the ABM component for warning about negative trends that are predicted in the system, or for suggesting necessary actions to neutralize negative trends, or even for deciding if some action in the system should be allowed or prohibited depending on its effect to the system. Social Activities Recommender (SAR) One of applications of pSNAs (proactive SNAs) in the AB is through the Social Activities Recommender component, a component that is capable of giving suggestions to users which actions should they participate in order to increase their efficiency or those that are more supportive of their own interests. SAR component is capable of learning or even asking users about their attitudes (objective goals) and shifting their behaviors toward their attitudes. We are implementing here results of research in the field of social psychology here. Namely, in social psychology, the cognitive dissonance occurs when individuals holds a belief/attitude but performs a contradictory action or reaction, i.e. it can be considered as an attitude-behavior inconsistency (Festinger, 1962). Through our SAR approach we are trying to reduce this inconsistency by promoting those actions to individuals and groups that will lead to activity-behavior consistency; but even further we are intentionally provoking this inconsistency, making individuals/groups aware of it as a first step toward more healthy behaviors. Group Leadership management. Breaking Hubs (Anti-preferential Attachment) Group leaders and motivators are basic motor of each group, the one that keeps it consistent and vibrant. They create events, introduce new activities, motivate groups, manage them, etc. Yet, different, even opposite anomalies can occur in this role, that we want to prevent, to control and to manage through AB. System identifies primary/actual group leaders and secondary/potential group leaders (Pearce 2002) and in different scenarios system manages them. We first describe different scenarios and later we introduce solutions for these scenarios. Scenario 1 - Totalitarianism - This is noticed through measuring group happiness (explained later) or through NLP tools. Lack of Group Polarization or Groupthink are other aspects of these scenario. Scenario 2 - No Leader - This is noticed through measuring group happiness (explained later) or through NLP tools. Lack of Group Polarization or Groupthink are other aspects of these scenario. Scenario 3 - Leader retires - Leaders may become inactive during some periods of time or may retire without announcement. AB detects that scenario and reacts before the consequences of scenario (lack of engagement) destabilizes the group. Solution 1 - Leaders promotion - AB motivates secondary/potential group leader to engage further. It suggests him to take actions, create events, AB promotes his actions and so on. This is activated in scenarios 1 and 3. Solution 2 - Auto pilot - By learning about group activities and defining groups patterns, system can autopilot a group, by creating events, suggesting already created events or new activities. Patterns learned among others are: how often members engage in a particular activity, which days are preferred by members for performing activities. This solution is activated in scenarios 2 and 3.. Solution 3 - Rotating leaders - AB introduces rotating team/group leader roles. By this approach, each group user is set by AB to be the leader for a specific period of time. This is activated in scenarios 1 and 2. Similar set of scenarios and solutions are identified and proposed for supporting small or new entities (buddies, events, venues, ) or for increasing global efficiency of the ecosystem (like number of people engaged in particular area, etc). In the most scenarios one can sense our tendency to neutralize negative aspects of preferential attachment phenomenon, or in other words our anti-preferential attachment tendency, which is one of examples of manually provided trans-technical patterns described by (social theory) domain experts (Fig 6).
HAPPINESS Happiness measures how happy actors are while performing activities with their peers (friends). More precisely, happiness is defined by measuring activities flow along actors relationships. We have modeled happiness as a level of balance of actions flow along two opposite directions of a friendship relationship . Therefore friendship that carry (transfer) balanced number of activities (gravitating to equal) in both direction is recognized as a happy friendship. The same applies to aggregated model of actor happiness, and system happiness. At the Fig. 8, one can see our ABM simulation of a system happiness in the response of the level of SAR suggestions followed by users. One desired and strongly necessary result was to get a strong increase of efficiency for a little sacrifice of users free will; we got over 60% of maximum increase of efficiency with only 20% of users activities according to SAR suggestions.
Fig. 8: ABM simulation of happiness. Each system incremented step of happiness correlates with system where users accepted higher level of SAR (FM) suggestions. Social Activities Recommendations based on Attractiveness Activity Buddies system introduces analogous relationships, that is, relationships that change in strength to represent different levels and phases in social relations, instead of binary yes/no relationships (Hays 1985). AB system introduces Attractiveness, which define how attractive is an entity (person, event, place, etc) to a user. Attractiveness is calculated based on several dynamic parameters, as shown below. System landscape is changing dynamically based on this, i.e. system notifications and suggestions are based on attractiveness. Use case: if person 1 lives at location A and has interests B and C, than system finds person 2 with same interests and same location to be attractive to her and notifies her about person 2 activities. But when person 1 travels to another place D, then system finds more attractive persons with same interests at location D and changes suggestions and notifications accordingly. Attractiveness is a complex value that is approximated in several of our ABM experiments, but we have identified the following parameters: 1) number of shared interests (in this calculation we consider only intersection of interests / IAmWhats between two agents), 2) mutual-friends - LOG N (filtered by: friendship.weight > threshold), 3) Traits, 4) how sociable (charismatic) is agent, 5) popularity - agents preference - preferential attachment degree, 6) male-female attractivity Attractiveness is used for calculation of very important measure for this model - attractiveness realization. Attractiveness realization Attractiveness realization measure says what proportion of user actions (as part of users TTPs) support increasing strength of relationships (friendships) that connect us with friends/agent attractive to us. It measures how much users, through the system usage, prefer attractive agents over non-attractive ones, i.e how well they realize attractiveness. In that way we are able to depict dissonance between our attitude (well-being) and our behaviour (Rabin 1994). Dissonance is the consequence of several different parameters, preferential attachment and environmental parameters (like geographical location) among others. Attractiveness realization (Fig 9) have values from 0 (minimal) to 1 (maximal) It is calculated for each agent as mean of functions of attractiveness and relationships strengths, for all of his neighbors. strengths-normalized = relations strengths toward all neighbors are normalized to range [0,1] attractiveness-normalized = attractiveness of all neighbors is normalized to range [0,1] attractiveness-realization is calculated for each neighbor relationship as following: attractiveness-realization = ratio among attractiveness-normalized and strengths-normalized telling us how much they deviate each from other
Legend: s - normalized relationship strength. Attr - Attractiveness; Attr-n: Normalized Attractiveness. Right image shows the case where is the biggest deviation between relationship strength and corresponding attractiveness, while left show the most correlated one. Fig 9: Attractiveness realization example Ad Hoc networking Above described dynamic functionalities enable ad-hoc networking. Compared to existing approaches, AB allows creation of interest frameworks in advance, by defining which activities (which interest) at what geographical areas during what period of time system should notify us about. This module enables active travellers to manage activities with persons on target destination. Our research has shown that the biggest problem in activities organizing is not being able to match with existing friends. In the same time our research shows that people are willing to perform (sport) activities with already unknown people. Built upon these two insights, we improve groups/networks scalability and flexibility through matching active travellers and people outside of the group to fill missing spots in group events. We additionally promote ad-hoc networking and weak ties through mediating in networking travelers with friends of their friends that live at travel destination and share desired interests. Choosing events based on attendees interests Activity Buddies covers need for networking at events, among peers who share or possess specific interests. Events range from scientific conferences till parties. Namely, besides coming for enjoying in events main content, one of the attendees key benefits is networking and experience with other people at the same event which will either at that event or some later event lead to more productive activities (Jehn 1997). Thus, we are enabling people to see statistics of attendees interests and to find and choose the event based on those data. CONCLUSION In this paper we presented an approach of utilizing advanced social analysis techniques in activity based online social network. We have promoted smart social systems, where system is redesigned, decomposed and layered in order to be more flexible to fit user behaviors and needs, but also in order to make it possible for system to learn social and behavioral patterns (or trans-technical patterns, TTPts) of social process (which we call, in their extended form, trans-technical processes, TTPs). Further, it makes possible for system to react in order to either fit system to user needs or to achieve system overall efficiency. This, redesigned type of system we call trans- technical systems (TTS) and the architecture that it conforms to, we call trans-technical system layered architecture (TTS-LA). TTPts in TTS dont need to be complex to achieve expected smart system goals. There is an important tendency to make recognizing, gathering, describing, implementing, and interpreting TTPts generalized and half-automatic. Fulfilling that goal will provide a cheap experts entrance to the system modeling and real-time interventions together with shifting system developers and designers from more monolith systems (from user perspective) toward SOA based systems, where components are dominating on all levels and phases of system production. Our future work will be focused in avoiding complex and non-unique grammar for describing TTPts, and in providing their integration through the system library/component and relatively simple set of visually guided TTPts produced by system/domain/global experts. More closely regarding the AB system and activity based social network, results of our research confirm that peoples attitudes dont match their behavior regarding active/healthy living. We also identify different socio- psychological factors and scenarios that are causing this attitude-behavior inconsistency. Accordingly, we introduce several approaches and tools that monitor, detect and react proactively in these scenarios. We show how (through different kind of mediations) an online system can improve offline world activities and prevent groups (inexorable) degenerative and dying out social processes. We show the way to increase amount of activities and to lead change of behaviors toward desired attitudes. After proving initial positive results through ABM modeling, in the parallel with running AB network on early adopters we will be able to test our concepts in-vivo and compare real-system efficiencies and benefits achieved utilizing them.
REFERENCES AND CITATIONS Wagner, G. (2000). Agent-oriented analysis and design of organizational information systems. Proc. of Fourth IEEE International Baltic Workshop on Databases and Information Systems, Vilnius (Lithuania). Taveter, K., & Wagner, G. (2001). Agent-oriented enterprise modeling based on business rules. Conceptual ModelingER 2001, 527-540. Sekiguchi, T., & Nakamaru, M. (2011). How inconsistency between attitude and behavior persists through cultural transmission. Journal of theoretical biology, 271(1), 124-135. Festinger, L. (1962). A theory of cognitive dissonance (Vol. 2). Stanford university press. Rabin, M. (1994). Cognitive dissonance and social change. Journal of Economic Behavior & Organization, 23(2), 177-194. Jehn, K. A., & Shah, P. P. (1997). Interpersonal relationships and task performance: An examination of mediation processes in friendship and acquaintance groups. Journal of Personality and Social Psychology, 72(4), 775. Hays, R. B. (1985). A longitudinal study of friendship development. Journal of Personality and Social Psychology, 48(4), 909. Chan, D. K., & Cheng, G. H. (2004). A comparison of offline and online friendship qualities at different stages of relationship development. Journal of Social and Personal Relationships, 21(3), 305-320. Cramton, C. D. (2001). The mutual knowledge problem and its consequences for dispersed collaboration. Organization science, 12(3), 346-371. Chiu, C., Hsu, M., & Wang, E. T. (2006). Understanding knowledge sharing in virtual communities: an integration of social capital and social cognitive theories. Decision support systems, 42(3), 1872-1888. Levin, D. Z., & Cross, R. (2004). The strength of weak ties you can trust: The mediating role of trust in effective knowledge transfer. Management science, 50(11), 1477-1490. Liu, C., & Tsai, C. (2008). An analysis of peer interaction patterns as discoursed by on-line small group problem- solving activity. Computers & Education, 50(3), 627-639. Pearce, C. L., & Conger, J. A. (2002, December 18). Shared leadership: Reframing the hows and whys of leadership. Sage.