Adaptive Social Networks - Research Paper Example

Contents Introduction 2 Background Information 2 Structure 3 Vector Space Model 4 Results 4Bootstrapping 5 Pivot Models 7 Stability 7 Conclusion 8 References 9 Adaptive Social Networks Introduction Conventional means of seeking information from the Internet, such as use of search engines like Google, Yahoo, and Bing among others or by getting information from databases are slower and often result in unreliable information. A major setback of using such means is that some information may not be available at the time the search was conducted. Another reason could be that the information is of sensitive nature so that the user is required to identify them before they can be allowed to access it. Background Information The issues to do with slowness, unreliability and insecurity of the conventional search engines motivated Bin Yu et al. to conduct an experiment on the development of a referral system. The referral system is to support interaction of agents with others by directing users to those agents who may provide the information they need. 1A research was done on an architecture that is fully distributed and has agents who maintain privacy and autonomy from its users. 2In 1997, Kautz and the group developed methods of graph analysis for referral systems where the person with the required information is found by use of referral chains. The informal person-to-person social networks are used to construct what is termed as the ‘referral chains of request. 3According to Bin Yu et al., the research conducted by Kautz et al. concentrated more on the problem of referral information generation at the expense of the dynamic aspects of referral systems. The aim of the research conducted by Bin Yu et al. was to study the dynamics of social structures which most referral systems studied. The agents in their architecture learn models of each other in terms of sociability and expertise. Structure They described an adaptive social network for information access that is intended to be an accurate, dynamic and evolving multi agent system that can achieve the effect of informal social networks that exist in an organization or community. In it, there is a distinction between a user’s interest and their expertise. Each user is associated with a personal agent. This agent is the first to see the user’s queries. He/she then decides the contacts to which the queries are sent. The agent that receives the query then decides if it suits its user and lets them see it. Alternatively, the receiving agent can respond with referrals to other users. The receiving user or agent can also choose not to respond in any way and discard the query. A query basically consists of a query vector and the requester’s ID, email address and a limit on the number of referrals requested. A response on the other hand consists of an answer, a referral or both of them. The answer given is highly dependent on the nature of the query and the level of confidence the referring agent has in the relevance of the referred agent. For successful interaction, the agent should know the user’s areas of expertise so as to decide whether or not to consent to their queries. They should also be equipped with knowledge on the models of their agents (neighbors) so as to decide whether or not to send them a particular query. Vector Space Model 4In their research, these authors adapted the vector space model (which is a classical information retrieval technique) to locate people and not the documents. The VSM estimates the importance of each term in a query and its corresponding power of discrimination amongst users to whom the query is sent. The user’s interests and expertise are represented as vectors in multidimensional information space. Then each query is systematically compared with the expertise vectors of other users to determine the one whose is most comparable to the query. The similarity between the two vectors is defined as the cosine of the angle between them. They normalized the expression by dividing the maximum length possible; the main idea being that an agent whose expertise for a query is as twice as large as another agent is similarly desirable as long as the angles of their expertise vectors are the same. “The query vector is generated from the user’s query while the expertise depends on each user considered as a potential target for the query”5. Because the expertise of the other user may be unknown, the querying user’s agent estimates the other user’s expertise learned from previous interactions. The computed similarity will then answer the query posed. Results Since each agent has an interest vector, an expertise vector and models of several neighbors, the neighbor’s model is dependent on the number of agents who know the particular agent, the number of agent the agent itself knows and the kinds of agents they are. In their architecture, the neighbor’s model kept by an agent is basically the agents’ representation of the other’s expertise and sociability. An agent’s queries are entirely generated on the basis of its interests. Once an agent receives a query it will either attempt to answer it based on its expertise or refer to other agents it knows6. The process is continued when the originating agent collects possible referrals and contacts some of them. This also changes the models of its neighbors. 7“In their experiment, Bin Yu et al. involved 20 and 60 agents with interest and expertise vectors”. The respective agents submit referrals, queries and responses to each other while learning about each other’s interests and expertise vectors. In their case, the agents were limited to 4 neighbors. These being the agents to whom a particular agent may send a query or issue a referral. Bootstrapping Bin Yu et al. began from “a regular ring but allowed for edges to be directed”8. They made use of a 100 node regular ring, and four edges per node as a start point for the experiment. “During every cycle of simulation, they randomly designated an agent to the “requester” agent and update the neighbor models of each agent upon the responses”9. Since they did not have real users, goodness of an answer depended on its closeness to the interest vector. Pivot Models In their architecture, pivots are modeled as agents who have a higher number of neighbors (out degree) than other agents; such agents are valuable to others because of their higher out degree10. This enables them to end up with a high in-degree as well. Their simulations confirmed the hypothesis that the existence of a pivot agent does enhance the resultant quality of the entire referral network as observed by all agents involved. When a pivot is introduced into a stable system, its quality improves significantly. This is because the pivot results in to more contacts that result into shorter paths between agents in various communities. Retrieval of the pivot from the system results in the opposite effect. Stability A referral network may not be eternally stable because agents keep on adding or removing themselves from the network. 11Bin Yu and the teams carried out what they termed as a “sensitivity analysis” of the social network. They did this by randomly introducing a new agent into a stable network. The new agent was randomly assigned neighbors but was expected to find those neighbors (gradually) whose expertise was closely related to their interests via referrals. Respective queries were only created by the new agent during the experiment. They found that the quality of the referral network improves as a result of the introduction of the new agent. Conclusion In their research, Bin Yu et al. did not come up with the best networks12. This is evident when a pivot is introduced into a network. The network’s quality improves significantly. However, the approach for referral networks is useful in not only aiding in faster and secure access to reliable information but also in building social networks for humans13. The research also shows that the quality of the networks improves when the clusters are reduced. Through referral systems, the user is not only able to find experts but helps them bring communities together in collaborative effort. References Kautz Henry, Selman Bart, and Shah Mehul. Referral web: Combining Social Networks and Collaborative Filtering. New Jersey: Murray Hill, 1997. Le-Shin Wu, Ruj Akavipat, Ana G. Maguitman & Filippo Menczer. Adaptive Peer to Peer Social Networks for Distributed Content Based Web Search, Department of Computer Science, School of Informatics, Bloomington: Indiana University, n.d. http://sixearch.org/paper/bookchapter.pdf. Yu Bin, Venkatraman Mahadevan, and Singh Munindar P. An Adaptive Social Network for Information Access: Theoretical and Experimental Results. Departments of Computer Science, Raleigh: North Carolina State University, 1997. Acessed April 16, 2012. http://www.cs.cmu.edu/~byu/papers/aai-referral-final.pdf. Read More