Research Projects

Polarization and formation of filter bubble in rec sys

We study the formation and evolution of filter bubbles (also known as echo chambers) in a recommendation system where content creators strategically adjust their content to their audience. Filter bubbles are undesirable artifacts of personalization in recommendation systems that expose users to content confirming their opinions. We study the problem theoretically under user-side and item-side fairness notions. We demonstrate that extreme item-side fairness constraints can lead to higher levels of polarization among users and expedite the formation of echo chambers. (Ongoing)

COVid-19 forecasting using deep learning

Proposing the use of deep learning to automatically learn the relationship between daily confirmed cases and viral load data. We trained one Deep Temporal Convolutional Networks (DeepTCN) and one Temporal Fusion Transformer (TFT) model to build a global forecasting model. We supplemented the daily confirmed cases data with viral load data and other socio-economic factors as covariates and compared the models' performances. Our results suggest that TFT outperforms DeepTCN and learns a better association between viral load and daily cases.

This paper was published at the "2023 IEEE International Conference on Big Data"

Human Activity Recognition

Implementing a hierarchical classifier to infer user's activities and capture different levels of abstraction in activity recognition. We proposed a two-level neural network classifier and showed that it can outperform the flat neural network with the same architecture. Our paper was accepted at the 12th international conference on Intelligent Human-Computer Interaction (IHCI 2020). Also, it is published in Springer Lecture Notes in Computer Science.

Price optimization in a network

In this project, a stochastic model of sales propagation among consumers is developed by specifying the transition rules between different states based on individual market parameters and the local network effects. Our goal is to maximize the profit by optimizing the prices for two problems of competitive and cooperative pricing. For each problem, we studied homogenous and heterogenous pricing. Our study develops an insight into strategic pricing for new products under local network effects.

University of Virginia COVID-19 Wastewater Surveillance

In this project, we utilize COVID-19 RNA concentration in wastewater (viral load) to predict the number of infections in the community. Viral load data contains information on pre-symptomatic, asymptomatic, and mildly symptomatic (PAMS) patients and signals an outbreak days before lagging indicators like confirmed cases, deaths, and hospitalizations. We made use of partially observed Markov processes (POMP aka HMM) and a compartmental model (SEIR) to simulate how COVID-19 spreads through the community.

We showed that viral load can accurately predict an outbreak of COVID-19 days in advance if used with a proper epidemiological model.

This paper was published at the "2021 IEEE International Conference on Big Data"
All the coding was done in R

Movie recommendation using implicit feedback

This research project focuses on enhancing the accuracy of personalized movie recommendations through collaborative filtering models, utilizing the MovieLens 1 Million (ML1M) dataset. The study compares baseline recommender models, including Matrix Factorization with Alternating Least Squares (ALS), KNN with Cosine Similarity for items, and KNN with BM25 Distance for items. Through meticulous preprocessing and training, the performances of the models were evaluated, and 95% CIs were computed for each evaluation metric. Furthermore, an ANOVA was applied to the MF model to ascertain the most important hyperparameter.

Implementing a sensor fusion algorithm

Implementing a new sensor fusion algorithm based on Dempster-Shafer (DS) theory of evidences. The implementation is done in Python and enables us to most effectively combine information from different sensors of smart devices to capture a more thorough and accurate of users' activities and states. The paper of this work is underway.

Matching Walking partners

This project aims to build a decision making framework for assigning walking partners intelligently. The objective is to create a robust framework that uses data from smartphones (and possibly fitness trackers) to assign walking partners to individuals. The algorithm is supposed to run automatically and frequently and suggest walking partners passively (based on their routines and their preferences).