Machine Learning and Natural Language Data Scientist

UPENDRA SAPKOTA

530-***-**** ***********@*****.*** 3014 Beltline Road, #2315 Garland, TX, 75044 EXECUTIVE SUMMARY

An award winning PhD graduate with skills in machine learning, unstructured data, and expertise in clustering, deep learning, supervised, semi-supervised, and domain adaptation models. Actively seeking Machine Learning and Natural Language Processing Data Scientist position. EDUCATION

PhD in Computer and Information Sciences, University of Alabama at Birmingham, AL (3.84) 2016

MS in Computer and information Sciences, University of Alabama at Birmingham, AL (3.83) 2011

B.E. in Computer Engineering, Tribhuvan University, Nepal (4.0) 2008 TECHNICAL STRENGTHS

Languages Perl, Java, C++, Python, SQL, Ajax, JavaScript Machine Learning Classi cation (Naive Bayes, SVM, Decision Tree), Deep Learning, Neural Net- work, Logistic Regression, Parsing, Ensembles, Part-of-Speech Tagging, Depen- dency Relations, Clustering, PCA, SVD, Domain-Adaptation, Character-based Models, Regularization

Protocols & APIs XML, SOAP

Tools SVN, Weka, OpenNLP, Stanford Parser, LibLinear, Cluto, Scikit-learn, Keras, Theano

EXPERIENCE

Freelance 02/2017 - 11/2017

Natural Language Processing and Machine Learning Researcher Dallas, TX

Computed frequent itemsets and association rules to help a local supermarket come up with better placement of items based on buyers preferences

Used Keras and Theano framework to build a model based on deep learning and identi ed the location of an object in given images

Used Scikit-Learn framework to explore Naive Bayes, and Support Vector Machine (linear and rbf kernel) to correctly classify employees to given incomes groups, and compared the performance with deep learning using Keras and Theano

Computed the clusters of similar groups using spectral and agglomerative clustering by representing the data with Scikit-Learn tf-idf vectors

Empirically veri ed the e cacy of n-gram categories in author pro ling

Explored di erent pro le-based methods on Query and Gender classi cation University of Alabama at Birmingham 05/2016 - 01/2017 Machine Learning Researcher Birmingham, AL

Explored di erent semi-supervised, unsupervised and transfer learning methods as a volunteer

Performed the usability of character-based machine learning models on cross-genre and cross-topic authorship attribution problem

University of Alabama at Birmingham 02/2010 - 01/2016 Developer/Machine Learning Researcher Birmingham, AL

Designed and developed an application to lter spam emails using Java

Created structured data from di erent types of unstructured data (twitter, emails, blogs, essays, reviews, news) and used feature extraction, feature selection, feature reduction (PCA, SVD) on it

Utilized centroid-based (k-means) and probabilistic (LDA) clustering models to create similar groups

Successfully used regularization to overcome under- tting/over- tting of the data, and cross-fold validation and resampling to validate di erent models

Demonstrated the e cacy of multi-modality models using similarity-based and pro le-based approaches that could be used for tasks involving data from di erent modality such as video, audio, and text

Designed a domain-adaptation method and demonstrated that our proposed new features ( pivot+new) has reduced dimension and is signi cantly better than traditional features (p=0.041) Yomari Inc. Pvt. Ltd. 08/2007-05/2009

Developer & Web Designer Kathmandu, Nepal

Used Java, SQL, JSP, JavaScript and Ajax to create student management system that manages data from large number of schools in a single platform

Developed server and client for a Stock Exchange Web Service & Investors Analyzing and Trading Project using SOAP, C#, java, and SQL Server

Created a shopping cart web application using Spring and Hibernate

Designed a grammar inference system using C++

PUBLICATIONS

Domain Adaptation for Authorship Attribution: Improved Structural Correspondence Learning. U Sapkota, T Solorio, M Montes-y-Gmez, S Bethard, ACL 2016

{ Demonstrated di erent domains have some corresponding features that can e ectively transfer knowledge from one domain to another, which is applicable to language translation, speech to text and text to speech problems

Not All Character N-grams Are Created Equal: A Study in Authorship Attribution. HLT-NAACL 2015

{ Empirically demonstrated punctuation category is signi cantly better than the word category with p < 0:01 (two-tailed t-test)

Improving the performance of cross-domain authorship attribution. U Sapkota PhD Thesis, The University of Alabama at Birmingham, 2015

{ Proposed and designed three highly e cient authorship attribution models using clustering, supervised, semi-supervised, and domain-adaptation models

Cross-Topic Authorship Attribution: Will Out-Of-Topic Data Help? U Sapkota, T Solorio, M Montes-y- Gmez, S Bethard, Paolo Rosso, COLING 2014

{ Proved addition of non-topic training data improves the performance by more than 50% HONORS/AWARDS

Outstanding Graduate (Doctorate) Student for scholastic achievement and leadership 2015

Full Scholarship and Fellowship from Government of Nepal in B.E. (2003-2008)

Semester topper award for three semesters in B.E. (2005-2007)

Contact this candidate