Research — Hemanth Bharatha Chakravarthy

Research

• The Economic Survey of India, 2021.

  in the 2021 Union Budget with Krisnamurthy V Subramanian and his Chief Economic Advisor's Office

• Chapter 2, Reports on Trends and Progress in Banking, 2020.

  at the Reserve Bank of India with Snehal Herwardkar (Director, DEPR) and Bhanu Pratap Singh
  ▸ Abstract
  The global banking system, bolstered by the progressive implementation of the Basel III reforms and swift policy measures, successfully withstood the initial impact of COVID-19. The implementation of further reforms was extended by a year to buttress the operational capacity of banks and supervisors to respond to the event. Going forward, the muted credit expansion, the persistence of a low interest rate environment and the impending asset stress on account of the pandemic suggest that profitability of banks is likely to remain subdued.

• Impact of the COVID-19 Lockdown on Indian Migrants and the Village Economy, 2020.

  authored the 2nd and 3rd round reports, as Research Associate to Heather Schofield
  ▸ Abstract
  We present a summary of the salient ways in which informal workers—who are disproportionately local migrants—have been affected by the COVID-19 pandemic and the associated lockdowns. This report is based on five rounds of household surveys conducted over the period of June - October. The sample consists of male migrant daily-wage laborers primarily employed in construction recruited in Chennai but mostly from villages or towns in Tamil Nadu, and their wives. Participants attended a minimum of one round and a maximum of five rounds.

• Variational Auto-Encoders (VAEs): A Review, 2022.

  Notes, benchmarks, explorations, and code examples
  ▸ Abstract
  “What I cannot create, I do not understand.” -Richard Feynman. I review the theory on why VAEs are relatively principled and interpretable vis a vis other deep latent-variable models. I suggest an implementation with torch and report benchmarks on MNIST and CIFAR datasets. I also explore the latent spaces of these toy models, and generate unseen (blurry) images by randomly sampling in the neighborhood of real encodings.

• Botocracy: BERT-powered Text Model for Predicting Authoritarian Twitter Platform Manipulation, 2022.

  with Em McGlone (Code available here, data available on request)
  ▸ Abstract
  As politics moves onto Twitter, authoritarians must update their set of tools used to manipulate discourse and amplify sentiments. While much attention has been given to automated “robots” that violate platform rules and post junk content, inadequate attentionis given to the paid armies of real users that strongmen leaders deploy to control trending charts and manipulate voter timelines. Our thesis is that the Indian paid armies of political Tweeters are exploited to add vitriol to the platform, drive nationalistic sentiments, and attack critics—that is, change the nature of the text corpus on Twitter. Thus, we are interested in predicting whether an account is a platform manipulator or not based on their Tweets' word embeddings and user metadata. An ideal dataset collected from scraping millions of Tweets during the peak of the 2019 Indian national election campaigns is used to subset-train the Google BERT base model and then test subsequent models built on the word embeddings. Through model tuning and evaluation, we arrive at a random forest natural language processing model for Twitter platform manipulation prediction. The model is trained on the most relevant principal components of Tweet word embeddings and metadata such as likes or days existed to predict a coordination indicator. The coordination indicator is predicted as true if the Tweet is associated with Tweets that were copy-pasted by multiple unique users. On the test set, the random forest model of choice has an accuracy of 74.07%, a sensitivity rate of 19.74%, and a specificity rate of 82.55%. This paper accompanies our NLP web application dashboard, a Twitter “botocrat” detector, that is built upon this random forest model.

• Lessons on Teacher Pay: Policy Brief on Performance-Based Pay for Teachers, 2020.

  Work in progress
  ▸ Abstract
  A comprehensive review of the literature on performance-based pay for public school teachers culminating in ideal mechanism design for pay-for-performance policies, with application to the Indian context.

• Evidence from the use of Twitter Bots by Modi's BJP in India, 2020.

  ▸ Abstract
  I explore the use of Twitter Bots in India by the ruling Bharatiya Janata Party and make multiple novel contributions to the study of bot use by the BJP. First, I offer rigorous evidence of bots reacting to key events with disaggregation of events and study of time lags showing that bot creation is highest when there is a dominating political event occurring and progressively falls with time or with lower national relevance of the event. Here I find a regression coefficient of 80.742***, between the bot creation in a day and key political events—evidence that more than twice the median number of bots created in a day are created on days with key events. This establishes the deliberate attempt to dominate political discourse around political issues perceived as important in public eye—more critical events result in more bot creation and as topics grow irrelevant with time, bot creation wears off to usual norms. Second, I present evidence around their strategy, clustering, and amplification: dual prongs of mass retweeting and mentioning “seed” bots and prominent pro-BJP sources. I describe seed bots, describe the bot network, its interlinkedness, and the existence of large clusters very quickly in the beginning of the tree diagram. Finally, I offer evidence of propagandic sentiments in use by Tweets and account descriptions and present quasi evidence of coordination between bot content. I also make contributions of public interest in understanding the prevalence of political bots in India and ways to identify them.

• Shiny Dashboard: Twitter in the World's Biggest Election, 2019.

  (Code available here)
  ▸ Abstract
  Studying Twitter’s response to the Main Bhi Chowkidar (I too am a guard) electoral campaign of the Bharatiya Janata Party (Background, Wiki), analyzing it using machine learning to study sentiments, representing views geospatially on a map, studying tweet frequency lumping and potential bot activity, and more.

• Shiny Dashboard: Gunshots in DC, 2018.

  with Ilkin Bayramli (Code available here)
  ▸ Abstract
  Interactive and animated geospatial analysis of Washington DC gunshot data: i.e., select your inputs to view animated maps of gunshots in DC in different time periods. A disclaimer: Shiny, while convenient, interactive, and open source (or maybe because of these) is terribly slow at loading interactive GIFs.