Mining Maximal Quasi Regular Patterns in Weighted Dynamic Networks

Under Supervision of Dr. Anand Gupta

Interactions appearing regularly in a network may be disturbed due to the presence of noise or random occurrence of events at some timestamps. Ignoring them may devoid us from having better understanding of the networks under consideration. Therefore, to solve this problem, researchers have attempted to find quasi/quasi-regular patterns in non-weighted dynamic networks. To the best of our knowledge, no work has been reported in mining such patterns in weighted dynamic networks. So, in this paper we present a novel method which mines maximal quasi regular patterns on structure (MQRPS) and maximal quasi regular patterns on weight (MQRPW) in weighted dynamic networks. Also, we have provided a relationship between MQRPW and MQRPS which facilitates in the running of the proposed method only once, even when both are required and thus leading to reduction in computation time. Further, the analysis of the patterns so obtained is done to gain a better insight into their nature using four parameters, viz. modularity, cliques, most commonly used centrality measures and intersection. Experiments on Enron-email and a synthetic dataset show that the proposed method with relationship and analysis is potentially useful to extract previously unknown vital information.

• Paper (Journal)

Framework for Mining Regular Patterns in Dynamic Networks

Under Supervision of Dr. Anand Gupta

Dynamic networks are categorised into four major types: Unweighted-Undirected, Weighted-Undirected, Unweighted-Directed and Weighted-Directed. Mining regular patterns in these dynamic networks finds extensive applications in prediction systems, analysing social network activity, etc. Depending on types of dynamic networks, regular patterns conforming to different characteristics of a system can be mined, each providing some unique information. Dynamic networks have pattern mining techniques that mine these characteristics individually. However, to the best of our knowledge, no such work is available that mines all these regular patterns (if possible) in a single run for any type of the dynamic networks. In this paper, the authors' focus is on providing a framework to make the task of mining regular patterns in dynamic networks more efficient and exhaustive. The accuracy of this framework has been verified mathematically and its efficiency is validated experimentally on a real-world network (Enron dataset).

• Paper (Journal)

Removing Noise from Dirty Documents using Machine Learning

Implemented a Kaggle Problem

Optical Character Recognition (OCR) is the process of getting type or handwritten documents into a digitised format.OCR makes previously static content editable, searchable, and much easier to share. But, a lot of documents eager for digitization are being held back. Coffee stains, faded sun spots, dog-eared pages, and lot of wrinkles are keeping some printed documents offline and in the past. As part of this project, I developed an algorithm making that uses Global thresholding to remove noise from images.

• GitHub

Conversion and Reconstruction of Compound Images to an Editable Form

Undergraduate Thesis Project

With the widespread digitisation of large document databases, the field of document image analysis has gained importance. There is a need to not only store these documents electronically , but also extract specific components of the document when required, and even edit these documents. While various OCR techniques are available to extract text from image, these techniques fail to recreate the original document in terms of its layout and text formatting. Also, documents today often contain more than just text which poses a further challenge. The additional components like tables and pictures also need to be extracted separately, and placed at their correct positions. In this thesis, we propose a model to extract the components (text, tables, pictures) from a compound document (containing one or more text blocks, pictures, and/or table components) and also, reconstruct the document in an editable format (LaTeX). Then, we experimentally demonstrate the working and verify the accuracy of this model. This project has been converted into a research article and has been submitted for publication in an International Journal.

Recognizing Human Activity Using Smartphones

The purpose of this project is to use the dataset provided at UCI ML Repository for detecting regular human activities. This dataset has been prepared from the recordings of 30 subjects performing activities of daily living (ADL) while carrying a waist-mounted smartphone with embedded inertial sensors. There were two ways of approaching the problem, either by training a user-dependent model which trains and tests a model using data from a single subject only in the dataset, whereas the user-independent model makes use of data from all the subjects. The former would have resulted in a higher accuracy but would have compromised on generalisation if a new subject is introduced, and therefore, I have gone ahead on implementing the user independent model. The proposed solution consists of an implementation of Support Vector Machine algorithm and the accuracy of the proposed solution turned out to be approximately 91% which can be further improved by using other algorithms.

• GitHub

Intent Parsing for Web Automation Tasks

Under Supervision of Dr. (Prof.) SK Dhurandher, Dr. Ankita Bansal

There exist several web automation platforms such as IFTTT and Zapier that allow power users to create automated workflows connecting together several services. However, these platforms, while incredibly powerful, fail to be able to leverage advances in natural language processing tools to build an interface that can be used naturally by novice users. My project is a natural language based interface for these frameworks. In particular, I am training a system making use of Google's Cloud Natural Language API and the dataset prepared by researchers at Carnegie Mellon University & Brown University to learn the intent behind relatively complicated natural language commands to make automated API calls and carry out actions on behalf of the users, as well as create programmatic rules for future behavior. This is much similar to Siri but for WebServices.

Dynamic Recognition of Tangible Object Placement Codes

Under Supervision of Dr. (Prof.) SK Dhurandher, Dr. Ankita Bansal

Northwestern and TUFTS HCI Labs has come up with a very high accuracy set of "Object Placement Codes" which are similar to QR codes but can be accurately identified over large distances and survive high variability in input. These codes can be used to tag real world objects to create customized triggers as desired. However, the current systems of recognizing these codes require that a picture be taken and then sent to a server, which then returns an output. There exists no web or mobile app based solution to firstly, do this in real time and secondly, cause software triggers to happen after recognition. Currently, my project focusses on the former while the latter is a strech goal for the same. Something similar to this is being done by Google (Project Bloks), and there are a large number of potential applications otherwise such as developing Mixed Reality Games, Tangible Programming Languages, Music Instruments etc.