Invited by Dr. SW Chen to Taiwan. This time work on a credit card data and phone logs collected from over a million customers (over 40 GB of Big Data!) Found relationships between financial troubles on Credit Card bills and call based metrics.
Solo trek
Climbed up ~4000m of the Indrahar-Pass trek solo over a period of two days. Got down as well.
Researcher at Rutgers University
Flew to USA to intern at Rutgers University with Dr. Vivek Singh on the Rutgers Wellbeing Study.Analysed Spatio-temporal social data captured via mobile phone over three months from 90 students. Found phone based behavioural markers (Phoneotypes) that can explain human cooperation with a very high accuracy.
Cultural Events
As the overall coordinator of the Quiz Club, led the IITK contingent to Nihilanth the inter IIM IIT quiz meet (first team to qualify in four years!) Had won several intra college quizzes.
Intro to Machine Learning
Did my first project in Machine Learning, trying to predict the use ofBikeShare in Washington Dc using past use and weather data. Achieved a global rank of 588 (current rank) out of ~4000 submission.
Blog for a Cause
Managed to write an article about cows which won Blog for a Cause, a nationwide social blogging contest.
Glorified Spelling Checker
Joined the Creative Cell of the Science and Technology Council core teamof IIT Kanpur. Manually corrected hundreds of grammatical errors to successfully publish the SnT magazine Scientia (no web version sorry). Also, served as the editor for Meander the literary magazine of IIT Kanpur.
Entrepreneurship
MyObjectify the first ever community cum Marketplace to make 3D printing more accessible in India. I worked on the web-end to create the initial 3D renderer for STL files, helped with the business plan, wrote marketing blogs and ordered lots of pizza. The startup featured among the 10 best start-ups at Disruptit, IIM Ahmedabad.
Pitch-ur-Product
Won the third prize in Pitch-ur-Product in E-summit'13, a nationwide Entrepreneurship Fest hosted at IIT Kanpur. The product was a pipe cleaning bot designed over the summers as the part of the Robotics Club at IIT Kanpur.I didn't tighten bolts, just wrote the pitch.
Student Mentor
Mentored three first year students for ESC101 which is a first level course in Computing where they're taught basics of C language. Thankfully they passed the course (with above average grades).
IIT Kanpur
Cleared the IIT Joint Entrance Exam (JEE) with a 99.3 percentile among 500,000+ students which allowed me to study in the prestigious IIT, Kanpur and live on one of the most beautiful campus in India as a Material Science student.
I am Born
Born in the City of Joy, then Calcutta now Kolkata.
Predicting Financial Trouble using Phone Call Data
I was invited by Dr Sheng-Wei "Kuan-Ta" Chen of Academia Sinica, Taiwan to conduct research in his Networks and Multimedia Labs. The project was to find behavioural markers that can indicate financial trouble for an individual.
Abstract
Financial outcomes for individuals has been traditionally determined by past payment history. Recently, researchers have used methods like time series and machine learning. Again, socio-economic status has also been inferred from mobile phone activity data and some even found interconnection between social and mobile features and spending behaviour. We decided to explore the possibility to predict financial trouble using phone activity data.
Methodology and Resuts
After doing a careful initial analysis and purging redundant users we ended with ~30,000 people who had trouble in at least one month over the financial year. The data however was naturally imbalanced so we had to use balancing techniques like SMOTE. We applied several classification techniques like GBM, SVM, Naïve Bayes, and created a model that can predict financial trouble in the upcoming 3 months using just 9 months of previous data with a ~0.75 AUCROC and 73 % Accuracy. We also found consistent results using different slots in the year. The metrics involved were a combination of both Call based and Transaction based features.
Conclusion
Thus, our work is unique as we diverge from the traditional credit rating techniques that rely heavily on past credit scores. We propose a hybrid method that can help people who do not have an established credit history get credit cards and eventually loans. Results published in PlosOne.
Why was this fun?
Working on over 40 gb of combined data was taxing and taught me patience, thankfully I had the support of a superfast workstation and talented colleagues. This also gave me a wonderful opportunity to explore and appreciate the amazing culture of Taiwan and ride the fastest elevator in the world installed at Taipei 101.
In the summers of 2015, Dr Vivek K Singh of Rutgers university gave me an opportunity to intern in his lab. The project was to understand why humans cooperate.
Abstract
Cooperation is a fundamental human concept studied across multiple social and biological disciplines. Traditional methods for eliciting an individual’s propensity to cooperate have included surveys and laboratory experiments and multiple such studies have connected an individual’s cooperation level with her social behavior. We describe anovel approach to model an individual’s cooperation level based on her phoneotype i.e. a composite of an individual’s traits as observable via a mobile phone. This phone sensing based method can potentially complement surveys, thus providing a cheaper, faster, automated method for generating insights into cooperation levels of users. Based on a 10-week field study involving 54 participants, we report that: (1) multiple phone-based signals were significantly associated with participant’s cooperation attitudes; and (2) combining phone-based signals yielded a predictive model with AUCROC of 0.945 that performed significantly better than acomparable demography-based model at predicting individual cooperation propensities. The results pave the wayfor individuals and organizations to identify more cooperative peers in personal, social, and commerce related settings.
Why this was fun?
This was the first time I had dedicated time wholly to research. The intern gave me perspective about choosing research as a career option. It highlighted the perks of conducting research, i.e. being able to do things at your own pace and read the topics you like. I tasted bitter rejection as the paper got rejected at a CHI conference and also felt the euphoria of acceptance at Ubicomp’16. Thus, writing the paper taught me perseverance and I learned about the importance of being able to deliver a convincing argument in written words
I hate it when websites play the same song again and again and again. So, my friend Saurav Prakash and I tried to create a better music recommendation system that recommends niche songs as often as the popular ones.We tested out system on the million song dataset.
Abstract
We study the problem of hubness in the high dimensional data spaces. Recently been attributed as a curse of dimensionality, hubness tends to make the nearest neighbour relations asymmetric. Due to hubness, certain objects tend to be present in the nearest neighbour lists of many other objects, but there are only a finite number of objects that are present in the nearest neighbour lists of the hub objects. As a result, the recommendation systems that are based on the similarity between songs recommend the hub elements repeatedly, while a large number of songs, ‘orphans’, are never recommended. Through proper scaling methods, the neighbour relations can be made more symmetric causing a reduction in the hubness.
Methodology
Conclusion
Overall, the best performance was obtained when the distance matrix was constructed using the Pearson distance measure and using Local Scaling—for the 10,000 songs dataset, the number ofsongs that were getting recommend was ~99%.
Why was this fun?
I love Spotify especially their Discover Weekly mix. So, I wanted to understand how recommender systems work and how I can improve one. Making the poster on Latex and presenting it to professors and peers was an enriching experience.
MyObjectify is the first online 3D printing community in India which aims to bridge the gap between talented designers and consumers so as to allow them to break away from the limitations posed by conventional manufacturing.
We believe that this can be accomplished by providing a platform where users can interact and iterate on the 3D design and then harness the power of additive manufacturing and rapid prototyping services, particularly 3D printing, to bring their 3D designs to life.
Why was this fun?
The two years I spent with MyObjectify were awesome. The layer by layer manufacturing technology never ceased to amaze me. The late night debugging sessions and qutiodian struggle of running a startup perfectly epitomizes my time at IIT Kanpur.
P.S. We also opened up a Mini ME stall at Mumbai ComicCon where we used a Kinect to 3D scan people’s heads and 3D print a miniature version of them.
Bicycle sharing programs have emerged as a global trend as an affordable, convenient, and sustainable travel option with various benefits. In this project, I tried different Machine Learning techniques on the Kaggle problem: Forecast use of a city bikeshare system. I used historical usage patterns with weather data in orderto forecast bike rental demand in the Capital Bikeshare program in Washington, DC. I did project as a part of the Machine Learning techniques course and this was my first project in ML. I learned how to do basic data analysis, do a feature selection and implement standard prediction techniques like Random Forest, Logistic Regression, Mixture Models and SVM.
Also, achieving a closing global rank of 588 (dropped 54 places) out of ~4000 entries for our submission on Kaggle was the cherry on the cake.
Almost at the end of my IIT Kanpur journey I decided to find the meaning of agood life under the mentorship of Dr Debayan Pakrashi of the Economics Department.
The aim of this project was to examine the markers of a good life in from of life satisfaction, relationship satisfaction, job satisfaction and satisfaction with children. Then try to see which variables affect these markers with a focus on heath and personality. I used the 2013 edition of the HILDA dataset which has an average of 17000 individuals collected over a period of 12 years. After identifying key socio-economic variables, I used a fixed effect regression model to find the importance of each variable. Many variables like job stress, # previous relationships, #children appeared to have a significant impact on life satisfaction while demographic variables like income, household type didn’t feature.
This project was completely different from the CS oriented projects I had done and introduced me to a different flavour of research and methodology.
Spatial Analysis of Crime in Uttar Pradesh
Intrigued by the high crime rate in Uttar pradesh, I pursue a data centric approach to find spatial connections on a district level.
Abstract
Uttar Pradesh (UP), the most populous state in India has been historically famous for its highcrime rate and lawlessness. In a 2012 survey, UP had been termed as the one of the "worststates" in India in terms of law and order by the National Crime Records Bureau (NCRB). UPis also notorious for its illegal arms trafficking and several illegal arms factories exist. I aim to findthe determinants of crime in UP and explore whether spatial patterns exist in the distribution ofcrime over the state using a Spatial Model for cross-sectional data and thus highlight major crime "hotspots" in terms of regional proximity.
Methodology
Conclusion
After applying the various tests I concluded that there is low spatial correlation withrespect to crime in UP. The crime rate is aptly explained by the socio-economic variables like unemployment and literacy in the OLS but since data collection isonly possible in urban cities we might not get the entire picture of crime in UP.
TA 101 was a first level Engineering Drawing course which was compulsory for all first year students at IITK. Armed with cumbersome mini-drafters, we had to draw orthographic projections of ellipses across sheets of white paper. Needless to say, the whole exercise was enervating. So, during the summers of 2013 I decided to unburden the freshers by designing a cool new digital drafter as part of the Des633A , New Product design course.
This introduced me to the heuristics of a new product design form ideation to prototyping. I learned about Quality Function Deployment (QFD) method to model the product according to user needs and the doing a STP to market the said product.
We even fabricated (and destroyed) a cool prototype for our drafter using an aurdino.
Echo: Educational app to learn Better Pronunciation
Echo an educational app to help children master English faster.It randomly chooses a word from database and displays it in the form of flash cards on the screen.It allows users to record their pronunciation which is assessed using tools like Google Speech2Text API and Chromaprint library for audio fingerprinting. The user can also listen to the correct pronunciation of that word from our database directly from the web app interface.
This was our submission for the Eduvate App-a-thon.
The aim of this project was to test the paper PERCEPTUAL ACCOUNT OF SYMBOLIC REASON, by Landy et al, which propsed that symbols may act as targets for powerful perceptual and sensorimotor systemsI built a JavaScript application to replicate the experiments proposed in the paper and tested it on 20 first year students of IIT Kanpur. The subjects were asked to do simple calculations like
8=2+3*y
Their response and response time was recorded. No personal data was collected.
We found some consistency with the original paper and our work was appreciated by Dr Landy himself.
This application analyse the moods of a person over a certain period of time based on her updates on any social media platform. It will then plot a graph of the person’s mood swings against time (the graph was made using YUI). It was also used to suggest music and images to a person according to his/her mood in the Yahoo HackU competition.
This was made during a 24 hour hackathon by four sophomores high on Redbull.
This was my “hello world” project at IIT Kanpur. People often reduce words to their phonetic equivalent non-dictionary words.So book becomes buk, and great becomes gr8. We tried to build a phonetic spell checker to correct such typographical errors, often made during texting.
We used the Metaphone algorithm to generate a list of close sounding words and then used the python NLTK library to implement N gram to find the closest dictionary words and suggest it.
P.S. I thought Python was a snake before embarking on this one.
P.S. We also opened up a Mini ME stall at Mumbai ComicCon where we used a Kinect to 3D scan people’s heads and 3D print a miniature version of them.
