I am currently working on the data-wrangling part of AceMap System. I used to work in various topics including Machine Learning & Data Mining Systems, Vehicle System and Website Constructions.
In this project, I studied the accent classification problem that given a recording of one person speaking a known script of English words, how could we predict that speaker’s native language. I hacked this problem through four major steps — word segmentation, feature extraction, clip classification and recording classification. Using MFCCs as feature vectors and Logistic Regression as classification method, I build a system to achieve an overall accuracy of 97%.
In this project, I discussed three models to deal with PM2.5 prediction problem, including basic Autoregressive Moving Average (ARMA) Model, Stochastic Volatility (SV) Model and Stock-Watson (SW) Model. Furthermore, I designed a new model combing SW Model with Time Series Neural Network and analyze its performance with above-mentioned models. Results show that this newly-designed model provide more accurate prediction of PM2.5 concentration levels for the next six hours.
|I designed a Traffic Signal Schedule Inference Model in CityDrive, a map-generating and speed-optimizing driving system. Through the crowdsourcing records of green-pass and red-stop at traffic lights, the system dynamically constructs a city map and then provides speed suggestion to the drivers in order to avoid red-stops. I implemented this algorithm and conducted simulation in Matlab. Results shown that the continuous speed advisory service effectively smooths traffic flow and significantly reduces energy consumption.|
Following are several selected websites: