Projects
A large body of research on social and information networks are pivoted around homogeneous entities within the network. Mining these networks have resulted in many interesting observations and applications, such as community detection, information propagation. However, most real-world networks are inherently heterogeneous comprising of different types of entities and relations. For example, in an e-commerce network, entities could be product, user, seller etc. If we model such a graph as a homogeneous network, we may loose important information around the interplay of those many entities and relations. Therefore, modeling a heterogeneous information networks (HIN) is imperative to capture the essential semantics of the real-world. To that end, we are currently working on a framework/methodology for generating recommendations in HIN domains that take the various interaction patterns among entity types and their attributes into consideration. We are also working on generating personalized recommendation in this context that take user preferences into consideration in a HIN.
Harnessing Object Linking for Efficient Retrieval (MS Thesis)
Machine Learning algorithms induce models from feature representations of the un-
derlying complex real-world data such as images, videos or natural language. There is,
however, a large class of problems where appropriate feature-value representations are
hard to arrive at. At the current state of the art, it is hard, for example, to classify a
movie as romantic or otherwise based on its content. “Romance" is an emergent notion,
that arises out of nonlinear interactions of the pixels in each frame, the frames with
each other, and the movie with the mental model of the observer. Collaborative recom-
mender systems cleverly bypass this hard problem by exploiting the knowledge of how
users rate items. Search and recommender systems benefit from effective integration of
two different kinds of knowledge. The first is introspective knowledge, typically avail-
able in feature-theoretic representations of objects. The second is external knowledge,
which could be obtained from how users rate (or annotate) items, or collaborate over a
social network. With external knowledge sources like social networks increasingly be-
coming pivotal in large classes of search and recommender problems, we envisage that
object relations can be exploited in several interesting ways that would complement the
traditional feature-theoretic models. This also has the favorable effect of significantly
reducing human effort involved in engineering an appropriate set of features tailored to
a task.
We propose a graph based architecture for principled integration of
traditional feature based representation and external knowledge. The architecture is primarily a network of information units. In order to empirically evaluate our approach, we
restrict the scope to text classification tasks and a book recommendation task. Our
experiments show a significantly improved classification effectiveness on hard datasets,
where feature value representations, on their own, are inadequate in discriminating between classes.
Modeled an agent that tries to reach out to the target nodes, but can only do so by increasing the probability of influencing nodes it already has some control over. We study a situation where an agent wants to reach out to target nodes, but can only do so by increasing the probability of influencing people he already has some control over. We formulate this problem and provide an easily implementable solution with theoretical guarantees. We believe that this framework is more directly applicable in networks like LinkedIn where entrepreneurs spend a significant amount of effort networking to ultimately target some big organization or a venture capitalist.
Humor DetectionDeveloped a system that can classify text as humorous and non-humorous. Humor-stylistic features and polarity of jokes were used to build the model. We scoped this task only to one liner jokes.
Spell CheckerBuilt a spell checker for word, phrases and sentences. Used context words and collocations to generate and rank suggestions. It was part of NLP course project.
Signature VerificationDeveloped a system to verify and authenticate offline signatures using Artificial Neural Networks. We used geometrical features of the pre-processed signatures to train the Neural Network.