Rahul Gupta

Experience

• 

  Inria

  Jun 2022 - Aug 2022

  Summer Research Intern

  Time-Aware Neighbor Joining | On-site Research Internship - Conceptualized a more generalized version of Neighbor-Joining method for creating phylogenetic trees from a distance matrix- Retained NJ's time efficiency and checked its mathematical properties followed by accuracy on real-world tests- Used python OOPs, providing modularity and enabling easy development and testing of different methods- Suggested use of Place-holder leaf nodes at distance 0 from the parent and usage of hashtable to store the labels of the nodes and labels of their leaves and their separation, making the code uniform, logical and highly efficient- Optimized the code to run a 5000X5000 matrix from 12 hrs to ~1hr by vectorizing operations using Numpy and using hash-tables- Tested the algorithm by creating toy-models covering various edge-cases and proceeded to test it on real-world data- Successfully observed clustering of sequences of closer in dates and also older sequences as internal nodes. Show less

• 

  Mastercard AI Garage

  May 2023 - Jul 2023

  Data Science Intern

  Temporal Boosting | AI Research - Conferred with PPO- Came up with a solution for dealing with concept drift by implementing Incremental Learning in an XGBoost architecture.- Achieved a efficiency of 20 times faster and 4-5 percent increase in auc-roc score versus retrained model on temporal datasets- Suggested and implemented use of STAGGER generator to synthesize custom concept drifting dataset, valuable for comprehension- Presented paper on Unified Bias Variance Decomposition about generalization for decomposition of constrained losses .- Implemented custom Bias Variance Decomposition of 0-1 Loss to get instrumental insight into the performance increase.- Undertook comprehensive literature surveys for improving performance, from Neural Attention Forests(NAF) to merge sort optimisation for finding best split for boosters, along with different sampling techniques. - Implemented a custom change in XGBoost architecture based model which took advantage of recency of new data for improvement Show less