Realtime Social Sentiment Analysis App (ongoing)

➢ Got geo-tagged tweets in real time by using the Twitter Streaming APIs

➢ Evaluated the sentiment of tweets based on Spouts-Bolts topology pipeline in Storm

➢ Stored these results in HBase, MongoDB, Redis and MySQL for different purposes

➢ Visualized real-time statistical results on a website based on Spring + Spring MVC

Personalizing Yelp Star Rating (2016) Detail

➢ Mining quality phrases from Massive dataset by using NumPy, Pandas, NLTK and R

➢ Reduced recommendation time by 95% by using raw RDD caching in Spark Cluster

➢ Increased matching F1 score by 5% by implementing a Belief Propagation algorithm

Database System for Madison Foodie (2015) Detail

➢ Implemented a 2000+ restaurants searching database system by using Oracle MySQL

➢ Developed Webpages for queries, reports and administration by using Java and JDBC

ELT & Data Matching for eBooks & iTunes (2015) Detail

➢ Built an optimized Python crawler with high scalability and maintainability by Scrapy

➢ Data cleaning, integration, reduction and transforming by using TF/DF, Jaccard Index

➢ Boosted the precision from 75% to 98% while improving the recall from 66% to 92%

Traffic Congestion Classification (2015) Detail

➢ Precisions using support vector machine with different kernel functions and neural network with different hidden units and output functions are compared by 3-fold cross validation

➢ Bagging (Bootstrap Aggregation) algorithm is introduced with the SVM and Neural Network chosen by first step, and confusion matrix among three different congestion levels is used to evaluate algorithms

➢ Investigate how predictive accuracy varies as a function of training-set size.Experiments show that ensemble methods outperform single classifiers, and Bagging-SVM with RBF kernel outperforms all other classifiers for this classification task

Gaze Pattern Detection (2014) Detail

➢ Mounting mobile eye tracking platform to monitor student attention and improve it when student is distracted.

➢ Adopting high-order polynomial regression to match pupil coordinate and target coordinate. We also use running average to get robust estimation of focus and recapture their diminishing attention by none-verbal cue

➢ Finding shows this platform can significantly improve students learning performance by 48% in the experiment