SRAG: RAG with Structured Data Improves Vector Retrieval
We propose SRAG, which adds structured data to chunks before they are embedded. The query also is augmented by structured data, like semantic tags, KG triples etc.
Results show that doing this improves the results by 30%. No changes to the infrastructure are needed. It is a simple method, but works very effectively.
Causal Effects of ESG Risk Scores on Stock Prices Using Instrumental Variables
This paper studies the causal effects of ESG risk scores on yearly stock market returns.
The results show that as the risk increases, the causal effects in a 2-stage least squares model are negative.
The yearly stock returns decrease as the environmental, social and governance risk increases.
Multi-Document Financial Question Answering using LLMs
This work explores financial question answering using knowledge graphs, semantic tagging and semantic RAG. The
results show that financial question answering can be improved using semantic tagging and knowledge graphs. The
financial questions used for evaluation are esoteric and difficult to answer, and the answers are not completely
obvious.
Topic
Driven Text Summarization with Defragmentation using LLMs
The paper describes a prompt based method to do text summarization by first planning to get a set of topics and
then generating a final summary from the topic summaries. The work can also be used to do conditional
summarization based on a given set of topics.
JCOL: A Java package for solving the graph coloring
problem (Journal of Open Source Software, 2019)
JCOL is a Java graph coloring package that does a cascade of three randomized graph coloring algorithms including
DSATUR, randomized Iterated Greedy and Min-Conflicts local search. Results show that it is able to outperform
ColPack on some DIMACS instances.
GKNAP: A Java and C++ package for solving the
multidimensional knapsack problem (Journal of Open Source Software, 2019)
GKNAP is a package to solve the multiconstraint (multidimensional) knapsack problem. It is implemented in both
Java and C++. The algorithm uses Lagrangian multipliers as constraint weights in a greedy algorithm. GKNAP is able
to solve almost all benchmarks to optimality.
Model
Adaptation via Model Interpolation and Boosting for Web Search Ranking (EMNLP, 2009)
By interpolating between several models (from different domains, but the same target), the algorithm creates a
weighted ensemble of ranking algorithms which outperform each individual algorithm for search engine ranking.
Three
Proofs that the Square Root of 2 Is Irrational
This short article surveys three proofs that the square root of 2 is irrational. The first proof is a simple proof
by contradiction and the other two proofs use field theory from abstract algebra.
Comparison of Stochastic Forecasting Models
This paper explores five stochastic forecasting methods like stochastic differential equations, arima, kalman
filter, and the bayesian filter. It tries to forecast stock prices of Microsoft, Tesla and Target, 54 days into
the future, and finds interesting results.
Comparison of Three Recent Personalization
Algorithms
This article compares three recent personalization algorithms including Taxonomy Discovery for Personalized
Recommendations, Multi-Matrix Factorization and Bayesian Personalized Ranking. It computes the hit rate @10 for
all three algorithms on a sample of 1 million users and 200 thousand items.
Hebbian Graph Embeddings
Using an error-free (derivative-free) associative learning update rule (Gaussian perturbed, as in annealing), this
algorithm learns the embeddings of the nodes (vertices) of a graph. Using nearest neighbors, the algorithm is
applied to link prediction, reconstruction and to a recommender system at Target Corporation. The algorithm
outperforms almost all current algorithms (except SEAL) in the task of link prediction and reconstruction.
Genetic Algorithm for a class of Knapsack Problems
A genetic algorithm is applied to the task of solving hard knapsack problems introduced by D. Pisinger called
Spanner knapsack instances. The algorithm is able to solve all Spanner instances to optimality, and so, it can be
applied to larger and harder instances for which the optimum solution is unknown.
Analysis of Greenhouse Gases
Using counterfactuals (through extrapolation), this work tries to predict the temperature increase in the
atmosphere when the CO2 and CH4 levels increase.
Randomized
heuristic for the maximum clique problem
A fast random search algorithm for the maximum clique problem. The algorithm finds good large cliques in DIMACS
benchmarks very fast, and these cliques can be used in further applications, like graph coloring or recommender
systems.
Patents:
Link prediction using hebbian graph
embeddings (Issued, 2022)
A patent that uses the hebbian graph embeddings algorithm for link prediction. Hebbian graph embeddings learns the
embeddings of the nodes in a graph and uses these embeddings for link prediction and recommendations.
Similarity learning-based device attribution
(Issued, 2021)
Using the probabilities outputted from a Gaussian Mixture Model as features, this patented algorithm uses a random
forest classifier for cross device cookie matching.
Markowitz
Portfolio Optimization
This is a course project on Markowitz Portfolio Optimization as part of a masters degree course. Portfolio
Optimization aims to generate proportions of the stocks to invest into.
A
One Pager on Bayesian Statistics in Baseball
This short one page tutorial explains how Bayesian statistics and hierarchical models can be used for analyzing
the batting hits of baseball players.
Introduction to Matrix Factorization for Recommender Systems
This short expository tutorial on matrix factorization introduces the reader to PCA, SVD and general principles of
matrix factorization for recommender systems including describing briefly on how to parallelize the algorithm to
run on potentially millions of dimensions.
Notes on Icebreaker
This is an expository tutorial on Icebreaker, a new publication from Microsoft Research, which is able to achieve
state of the art performance on tasks in which there is inherent missing data. Icebreaker is able to suggest which
attributes to impute to gain the maximum benefit, using mutual information. The tutorial talks about the Bayes
Rule, Autoencoders, the evidence lower bound, Variational Methods and Autoencoders, Deep Sets (set based machine
learning) and finally leading to Icebreaker.
Analysis of
E-Commerce Product Graphs
This work analyses the e-commerce product graph at Target Corporation using the Erdos-Renyi random graph model.
Results show that the graph exhibits assortive mixing, the small world property, and a scale free degree
distribution (power law).
Open Source Code:
Taxonomy for Recommender Systems (Python)
Implementation of the methods in the paper Taxonomy Discovery for Personalized Recommendation. The algorithm
alternates between path sampling (Gibbs sampling) and latent factor updates (stochastic gradient descent).
Bayesian Personalized Ranking (BPR) (Python)
Bayesian Personalized Ranking (BPR) is a recommender systems algorithm that can be used for personalization in
movie rental services, retail stores, online book stores and so on. This is a python implementation which achieves
51% test set hit rate @ 10 on the MovieLens data set.
Genetic Algorithm for the Maximum Clique Problem (C++)
A genetic algorithm for the maximum clique problem that uses innovative crossovers. The algorithm is able to solve
publicly available instances quite efficiently. Results are on the git page.
Heuristic for Graph Coloring (Java)
A very powerful Java implementation of a heuristic that requires very little computational resources. The code is
able to solve several of the DIMACS instances in a short amount of time.
Sudoku Solvers and Puzzle Generator (C++ and Java)
A backtracking Sudoku solver which is non-recursive. It is inspired by Knuth's dancing links algorithm. The
backtracking solver is quite fast and solves most Sudokus in less than a second. A Java based puzzle generator
generates Sudoku puzzles which have a unique solution.
Iterative Local Search - Quadratic Assignment Problem
(C++)
The QAP (quadratic assignment problem) is a strongly NP-hard problem which is difficult to solve even for very
small instances. This algorithm is able to solve the nugXX (e.g. nug22) instances quite fast. More experiments are
needed to see if the algorithm scales to very large instances.
Branch and Bound Algorithm for the 0/1 Knapsack Problem
(Java)
Branch and bound is an exact algorithm that can solve the knapsack problem with provable optimality. This Java
code implements branch and bound for the 0/1 knapsack problem using Lagrangian relaxation to prune the search
tree. It uses best first search and is able to solve moderately large instances quite fast (like 2000 variables).
Monty Hall Simulation (Java)
A simulation of the Monty hall paradox in probability theory.
Education:
M.S. Applied and Computational Mathematics, Johns Hopkins University
M.S. Computer Science, University of Southern California
Computational Linguistics, University of Washington
