Social Media Analytics
(July 2017 - Dec 2019)
Worked with the Windows Social Media Analytics team to classify User feedback data from Social Media Channels and Feedhub Hub for better bug triage and automated actionable next steps. Leveraged Machine Learning and Natural Language Processing Techniques to conduct various experimentations for feature selection, tuning parameters and model training for picking the best parameters for production. Also did experimentations around merging structured and unstructured data to create a better balanced dataset.
(November 2016 - present)
Led a team to build a Fake news detecting Chrome extension "Project Fib", Google Moonshot Prize winner at HackPrinceton 2016 Featured on Business Insider, Washington Post, The Next Web, Mashable, Huffington Post, Microsoft Blog, Wired, Hacker News, Fortune, CNN, BBC, CBS, CBC News, Bloomberg TV, Boston Globe etc -
Invited to "Forbes 30 Under 30 Conference" 2017 for the work 50K User Requests per min, 650 stars on Github and 20K Dev views on Github page on 1st week
Won Grinspoon Entrepreneurial Concept Award 2017
Link - http://projectfib.azurewebsites.net/
Research Software Engineering Intern,
(July 2014 - August 2014)
Implemented an US state level Location Inferring Classifier for Twitter Users with overall 21% accuracy boost from Baseline for Project Trupil ( http://research.microsoft.com/en-us/projects/trupil/ ), lead by the Social Analytics Team.
This SVM Classifier, built with 5.1 million user tweets, uses the text content of the User's tweets (indicator of dialect of the region), time-of-tweeting as the main features and reverse geocoded latitude-longitude info of the location tagged tweets as class label
1. Masters Thesis : Detection of Fake News from Social Media.
Built Fake news detector Chrome Extension "FiB", which was featured in Washington Post, Business Insider, Mashable, Financial Times, TheNextWeb, Hacker news, The Huffington Post, the Boston Globe, Microsoft Blog, CBC News, CNBC News, CNN News, BBC News, CNET etc. - https://projectfib.azurewebsites.net
2. Analyzing Twitter User Profiles - It involves analyzing User's personality based on user's Twitter profile given his/her twitter id, by looking at the texts & image content of User tweets to do Sentiment Analysis, Tweets Topic extraction, Named Entity recognition, political views extraction and language detection.
3. Detection of Cyberbullying from Social Media using Machine Learning - Given a set of tweets, the classifier model can detect with 78% accuracy if its an instance of cyberbullying or not
4. Music Recommendor based on User's Mood from Social Networking Sites - Implemented a Music Recommendation system (Accuracy : 86%), which, given an input of User’s Twitter UserID, predicts the User’s mood as one of 20 mood classes from the User’s recent tweets using a Naïve Bayes Classifier & plays/recommends relevant mood song tracks mapped to Gaana.com. The training set of the classifier consists of 85% automatically annotated labels using mood hashtags tweets retrieval using Twitter4J & 15% manually annotated tweets and the classifier is built using 32 kinds of emoticons as well as alphabetical characters as features through Weka JAVA API.
Compared the performance of various Classification algorithms on a smaller training dataset.
This project’s paper won the IEEE Student Paper Contest 2015 for BPDC & is further submitted to compete in IEEE Student Paper Contest for R8 region.
This project was also selected as top Graduation Sponsorship Program 2014 projects by Microsoft Research, ATL Cairo.
5. Text Analysis From Social Networking Sites - Implemented a system that generates trending topics from current tweets through Twitter Streaming API. Performed Sentiment Analysis on tweets, retrieved by Twitter Rest API into positive, neutral and negative classes & also rating them on a scale of 1-5. Tools used : Stanford NLP, Alchemy API
6. Help Her - A chatbot on Social Media Platforms which helps domestic violence and attack acid victims to find more about laws related to their case. The platform utilizes Natural Language Processing and Machine Learning to figure out what case, this would fall under and further uses location information to find the nearest law centers.
7. MoodMusicBot - A Facebook Music Chatbot which maps Mood to Music
Given an Image, the chatbot can figure out the mood of the user using Microsoft Cognitive Services Emotion API and retrieves the appropriate mood based music from our Database and recommends songs.