Education

Doctor of Philosophy in Computer Science at Iowa State University, Ames, IA, US        (Fall 2017 - Expected graduation: May 2022)

• Adviser: Dr. Qi Li in Data and Knowledge Mining Group

Master of Science in Artificial Intelligence - Computer Engineering at Azad University of Tehran-Science and Research Branch, Tehran, Iran

• GPA: 4

• Dissertation topic: Design and Implementation of DECAPTCHA system (4 out of 4)

Bachelor of Science in Software Engineering - Computer Engineering at Azad University of Tehran-North Branch, Tehran, Iran

• GPA: 3.75

Publications

• [SDM'22] Adithya Kulkarni, Nasim Sabetpour, Alexey Markin, Oliver Eulenstein, and Qi Li. CPTAM: Constituency Parse Tree Aggregation Method. SDM 2022 proceedings, to appear, 2022.

• [ICDM21] Nasim Sabetpour, Adithya Kulkarni, Sihong Xie, and Qi Li. Truth Discovery in Sequence Labels from Crowds. Proc. of 2021 IEEE Int. Conf. on Data Mining (ICDM’19), to appear, 2021.

• [EMNLP20] Nasim Sabetpour, Adithya Kulkarni, and Qi Li. OptSLA: an Optimization-Based Approach for Sequential Label Aggregation. EMNLP'20 Findings, 2020. [Code] [Presentation] 

Teaching Assistant

• COM S 363 -Introduction to Database Systems

• COM S 574 -Machine Learning (Graduate Course) --Iowa State University, Department of statistics; Spring 2019 and 2020. 

• COM S 561 -Advanced Database (Graduate Course), with recitation class responsibility --Iowa State University, Department of Computer science; Fall 2018.

• COM S 311-Design and Analysis of Algorithms, with recitation class responsibility -- Iowa State University, Department of Computer Science; Fall 2017, Spring 2018. 

• Network and Security -Azad University of Tehran, Department of Technical and Engineering; Fall 2015.

• Peer tutor-- Design of Algorithm -Azad University of Tehran, Department of Technical and Engineering; Spring 2015.

Internships

• In the position of Data Science Intern; Boston, MA, US; May 2018 to Aug 2018.  

In this research, we present machine learning solutions as targeting strategies for the programs based on available alumni data (Behavioral, Analytics, Events, Gift history, RSVPs, etc.) in almost ten institutions. The Regression, Gaussian Naive Bayes, Random Forest, and Support Vector Machine algorithms are used and evaluated. The test results show that having been trained with enough samples, all four algorithms can distinguish donors from rejectors well, and more interested donors are identified more often than others.

The results show that in a practical scenario where the models are properly used as the targeting strategy, nearly 85% of new donors and more than 90% of new interested donors can be acquired. The test results show that all used algorithms can distinguish promising donors from non-promising donors. -Those who never upgrade their pledge. The best model produces an overall accuracy of 97% in the test dataset. 

• In the position of Data Intelligence Scientist; Boston, MA, US; May 2019 to Aug 2019. 

Finding a set of reference points which can be used by the meteorologists in their daily process to tune the demand forecasting model output has always been a challenge. This research presents a model that can hopefully lead us to the final goal of minimizing the differences between the actual demand and the nearest weather demand model output. Here, we construct a model to find highly similar weathers using historical weather data having attributes (Date, Temperature, Humidity, Humidity Index, Dew Point, Cloud Coverage, etc.), and it has the potential to be expanded to weather forecast model in the future. Here, we settle to use the Humidity, Temperature, Cloud Coverage and Dew Point as our model inputs.

This study focuses on providing a search algorithm that sifts through available historical data and selects the most similar historical dates based on multidimensional weather features. The K-Nearest Neighbors’ algorithm is based around the simple idea of predicting unknown values by matching them with the most similar known values. Then, to construct our “Almost KNN model” (a method which is very similar to K-Nearest Neighbors), we first construct the feature set from raw data, then compute the distance based on different similarity measures of Euclidean, Manhattan, and Cosine, and at the end pick the top k closest distances. In terms of time complexity, one challenge we encountered with Almost-KNN method is computational complexity. The amount of computation required for computing and sorting pairwise distance grows exponentially.

In order to provide an efficient and fast search time, we applied a K-Dimensional tree (KD_Tree) to partition our data into a more efficient structure for multidimensional searching. It is a kind of binary search tree to do space partitioning for organizing points in k-dimensional space. The results show that the combination of almost-KNN and KD tree can provide a computationally efficient algorithm for nearest weather day matching.