My research focus has been on mining data from web, social networks and make inferences on it. I have a keen interest in studying online social networks, both from content and network structure point of view. Here I try to capture some of my past and existing research threads that I am working on.
Studying the Impact of Rumors on Microblogs
Collaborators : Aditi Gupta and Dr. Ponnurangam Kumaraguru (IIIT-Delhi)
This work is done in collaboration with PreCog research group @ IIIT-Delhi. We collected data of two events namely Sandy Hurricane and Boston Marathon bomb blasts. We got the ground truth dataset of what rumors were being spread over this event, how they were spread and through whom were they spread on Twitter. The aim of the project lies in being able to characterize the diffusion process, susceptible users or classify the rumor tweets on Twitter. The Sandy work was accepted at Privacy and Security in Online Social Media workshop, at WWW, 2013 and the Boston work was published in eCrime Researchers Summit. The work has been gathering a lot of attention from the media as well. Here are some of the highlights:
http://irevolution.net/2013/10/20/fake-content-on-twitter-boston-marathon/
http://www.dailydot.com/news/twitter-boston-bombing-coverage-lies-spam/
http://voiceofrussia.com/2013_10_26/Boston-bombing-tweets-don-t-trust-Twitter-3333/
http://www.buzzfeed.com/rachelzarrell/most-of-the-tweets-during-the-boston-marathon-were-inaccurat
(and many more...)
Community Driven Supervised Topic Models for Prediction
Collaborators : Dr. Tanveer Faruquie, Danish Contractor, Kanika Narang (IBM-Research)
The goal of the project is to predict the movie revenues using the buzz generated on social media websites along with social influence information. We modified the supervised topic model based on the topics and sentiments of the posts. The algorithm also produces communities of users and distribution of topics and sentiments over it.
Signature of Graphs with Application to SNA Tasks
Collaborators : Dr. Ramasuri Narayanam and Dr. Dinesh Garg (IBM-Research)
The work is primarily focused on coming up with a signature of a particular network (which is less than the size of the original graph). The signature then can be used for any SNA task in particular. The way of generating the signature delves into the domains of graph sparsification. The metrics used for sparsification includes both structural and activity / dynamic content based features. We have shown the efficacy of approach with community detection and influence maximization as applications. Works related to this project has been accepted at WISE and COMSNETS.
Edge-Less Community Detection
Collaborators : Dr. Ramasuri Narayanam (IBM-Research)
The basic idea behind this particular thread is to come up with a community detection algorithm which does not take into account edges of the network while forming communities. An example of this is the work by Barbieri et Al "On Influence Based Network Oblivious Community Detection" accepted at ICDM, 2014. The aim is to not infer the edges from the cascade flow and then find communities but it is to find communities without inferring the edges. It is not necessary however that cascade information is available over all the datasets and only it should be used to find communities, any other data could also potentially serve as basis to find communities. A very very basic version of this idea is accepted at I-CARE, 2013 and a short paper was accepted (later, not published due to no registration) at WISE, 2013. An advanced version is still Work in Progress.
Study on Influence
Collaborators : Dr. Ramasuri Narayanam, Kuntal Dey, Seema Nagar, Natwar Modani (IBM-Research) and Dr. Ponnurangam Kumaraguru (IIIT-Delhi)
I have been interested in the area of influence maximization / what is influence? and related concepts like homophily and influence etc. since my undergraduate years (Aug 2011). I have been working on numerous problems on concepts related to influence. Currently, I am involved in threads involving data based influence maximization approach ( similar to Bonchi "Data-Based Approach to Social Influence Maximization") and quicker, customizable viral marketing approaches. Work related to this are under review and extended versions are still Work In Progress.
My undergraduate thesis work was also on finding influential posts in time of crisis and non-crisis. The goal was to be able to characterize if a tweet is more influential because of its content, author, time at which it is posted etc. It was also to find out the differences in the properties of tweets which were posted during crisis situations and non-crisis situations.
Like-Minded Communities
Collaborators : Natwar Modani, Dr. Amit Nanavati (IBM-Research)
Nowadays, for enriching customer experience or client - engagement or for launching intelligent viral marketing campaigns, it is essential that the communities found should not only take into account just the network structure but also the other information about each person that exists in the dataset. The communities thus founded add more meaning to themselves by not only defining the network structure of it but also identifying patterns as to why a person is in the community.
Susceptible Users In Social Media
Collaborators : Aditi Gupta , Dr. Ponnurangam Kumaraguru (IIIT-Delhi) and Dr. Ramasuri Narayanam (IBM Research)
Two of my projects have been on finding susceptible users on social media. These users can be either susceptible to influence or they can be susceptible to rumors on online social media. The goal in both these cases is to find such users. For susceptibility to influence, we used popular information propagation methods to determine who is susceptible and can we predict given a network which user will be susceptible using it's structural properties. For the other case of rumors / misinformation, we use the rumors in Boston marathon data and analyze what type of users fell for rumors and who do not. I have given a talk on this topic at EURO MMXIII (26th European Conference on Operational Research).