Rumi Ghosh

Contact information: USC Information Sciences Institute 4676 Admiralty Way Marina del Rey, CA 90292 tel: (310) 448-8302 email: rumig[at]usc.edu

I am in my second year in the Ph.D program in Computer Science at University of Southern California. I am currently working under the guidance of Professor Kristina Lerman at the USC Information Sciences Institute in the field of machine learning and information integration. I had completed my undergraduate studies at Indian Institute of Technology, Kharagpur from where I received a Bachelor of Science and Master of Science in Mathematics and Computing.

My research currently focusses on social information processing, augmented social cognition and collective intelligence i.e"wisdom of crowds".

I believe there exists an interdependence between people and smart machines; in the sense that smart machines help people and at the same time the study of inherent smartness of people can help building of these machines. Multiagent systems having correlation between intelligent agents comprising them, can make the human life easier in innumerable way; but at the same time development of such agents requires a closer understanding of groups working together. Naturally occuring intelligent groups about whom data is easily available and can be analysed are- people. This motivates the need to study social networks to build smarter machines.

The i.i.d(independent and identically distributed) assumption of random variables is common in statistics, probability and machine learning. Independent and identically distributed data can be viewed spatially as nodes of a graph with no edges between the nodes. But in case of real life data sets, the independence and identical distribution assumption does not often hold. Such data sets can then be represented graphically as nodes which are connected by edges giving rise to a network on which computational tools using i.i.d assumption no longer works. This demands the need for creation of tools independent of this assumption. One way to develop such mathematical and computational models is to analyse existing networks, extract their distinct characteristics and build tools using these properties. Such an analysis demands that the network under study should be easily observable. It should also be relatively easy to deduce its expected characteristics which would make it easier to cross validate of the results obtained. Social networks are easily observable and the rich literature existing in sociology, psychology, cognitive science, neuroscience can be used to deduce the expected charcteristics collective human behavior. This motivates the need to study social networks to build better computational, mathematical and analytical tools.

People are inherently good at data integration and semantic recognition. However, when the process of data integration and semantic recognition takes place at group level(leading to a hierarchy of information processing in which an individual is the lowest level), the errors occuring at lower levels slowly get eliminated with the increase of level leading to the gradual emergence of relatively accurate and semantically correct integrated information. This motivates the need of study of social networks to build better data management and information integration tools.

Study of social networks can also help in link prediction, trend prediction and detection of the influence hierarchy within the group. This in turn can be used in customisation of commerce and advertisements. It can also be used in determing the trends of the market and predicting the future trends, extracting hidden opportunities, building stronger collaborations, evolving better marketting strategies and in mastering social ecosystem marketing.

Last but not the least, analysis and social information processing helps in improvement of online social networks which can be powerful tools for staying in touch, extending connections and increasing visibility.

The projects that I am currently involved in:

• Automatic Community Detection and Hierarchy of Influence within Communities: We have deviced an algorithm for automatic detection of communities using a measure of global influence. It is generalization of the modularity based approach propounded by Newman et al. The topology of the graph primarily comprises of the the ties between the players i.e who is connected to whom.The algorithm computes the communities using the cascade model of influence spread, represented by an influence matrix.The metric used for the computation of influence is the total number of paths present from the former player to the later.The underlying hypothesis is that, more the number of paths from one player to another, more is the capacity to influence. The model depends on two parameters, the direct attenuation actor and the indirect attenuation factor, which basically are used to calculate the probability of transmission of influence from one node to another. Hence they model the strength of ties between the players. The details of the algorithm is available here.The algorithm has given us promising results in standard data sets like Zachary's Karate Club,College Football,Political Books and Southern Women data sets.We are presently implementing it on online social networks like Flickr and Digg. The hierarchy of influence of the people comprising the community, calculated using this method reduces to the Katz measure, which has been shown to be the most effective measure for link prediction.


• Microscopic Dynamics of Collective Behavior on the Social Web: The goal of the project is to understand the role of networks in spreading information. We will use agent-based models to study the microscopic dynamics of this process. We plan to conduct this study on the social news aggregator Digg, which allows users to submit links to and vote on news stories. The social networks influence much of the user activity on Digg and can be leveraged to gain insight into the collective intelligence of a networked community. The results of our study will be microscopic agent-based and mathematical models of collective voting.The microscopic models will provide new insights into the dynamics of collective decision making, including the incentives users have in participating on these systems.


• Individual and collective spatial representations in crowds and other social behaviors: Development and Study of Phenomenological Models of Collective Pedestrian Dynamics. Details available here .

I have been previously involved in:

Conceptual Binding Model of the Human Mirror Neuron System: I have been interested in understanding the working of the human brain and think some of the best machines have been built by replicating nature. Since my ultimate aim is to build intelligent analytical tools and machines, I undertook this conceptual brain modelling project under the guidance of Professor Michael.A.Arbib . The integrated conceptual neural model that we developed can possibly account for schizophrenia, autism and apraxia to a large extent. Details are available here.

Rumi Ghosh and Kristina Lerman (2009), Structure of Heterogeneous Networks, IEEE International Conference on Social Computing, Vancouver, Canada, 2009.

Rumi Ghosh and Kristina Lerman (2009), Leaders and Negotiators: An Influence-based Metric for Rank, International Conference on Weblogs and Social Media, San Jose, CA, 2009.

Rumi Ghosh and Kristina Lerman (2008), Community Detection using a Measure of Global Influence, Submitted to Springer Lecture Notes on CS, extended version of the paper at the 2nd Annual KDD Workshop on Social Network Analysis, 2008.

Dylan A. Shell, Shivakumar Viswanathan, Jin Huang, Rumi Ghosh, Jie Huang, Maja J Mataric, Kristina Lerman, and Robert Sekuler(2007), Spatial Behavior of Individuals and Groups: Preliminary Findings from a Museum Scenario , Proceedings, IEEE/IRSJ IROS 2007 Workshop From Sensors to Human Spatial Concepts, San Diego,USA, Oct 2007.

I enjoy painting, writing poems, theatre and learning anything that is new. I am currenty learning international ballroom dancing.
I also enjoy going on hikes and kayaking.

• Oct 1st - 4th: I attended the Grace Hopper Conference at Keystone, Colorado, thanks to a full scholarship from Northrop-Grumman and USC.
• Oct 3rd: Was runners up in Poster Presentation in the 3rd Annual ISD Graduate Student Symposium at USC/ISI. My poster was entitled Of Football and Ladies Who Lunch: What Networks can reveal?. A big thank you to my adviser, Kristina, for explaining my poster to others.