Melanie Ho

• R (ggplot2)

Established the predictive model for presaging the probability of having serious genetic disease based on personal blood test and background.

According to the raw data we had, utilized CART in filling missing values in blood tests reults (especially for blood test 3 and blood test 4), and applied principal component analysis to detect collinearity.

Click the pdf icon above, or scroll down to the file, or contact me to go deeper in this project!

• R (ggplot2)

Coming soon!

• Python 3.6 (numpy, tensorflow, keras, OpenCV, dlib, skimage)

Applied CNN and HOG+SVM in detecting human emotions and compared the accuracy.

By using CNN, the resulting model with two convolutional layers and full landmarks were processed in keras by inputting images into the convolutional neural network and guiding CNN by normalized landmarks.

By using HOG+SVM, detected facial features by extracting the histogram of oriented gradients feature descriptors and classified by using support vector machine.

Click the pdf icon above, or scroll down to the file, or contact me to go deeper in this project! It's github is coming soon!

• Python 3.6
• Anaconda
• Apache Spark 2.2.1
• Pyspark 2.2.1

Implemented a basic word counter in Apache Spark and resulted in a customized word-count dictionary.

According to the books inputted, the word counter will detect the words, record the corresponding counts, elminate the stopwords, filter out the punctuations, and provide the top n most frequent words in a dictionary. Otherwise, depends on the TF-IDF of each specific word, the counter will also produce the top t words with highest TF-IDF scores.

Click the github icon above or contact me to go deeper in this project!

• Python 3.6
• Anaconda
• Apache Spark 2.2.1
• Pyspark 2.2.1
• Google Cloud Platform

Implemented Naive Bayes classifier on Apache Spark and resulted in 94.52% classification accuracy in large dataset.

The project uses Reuters Corpus but only specifies four categories that end with CAT. We applied several preprocessing skills like removing punctuations, words stemming, and removing stopwords. The Naive Bayes classifier is also implemented with Laplace smoothing to avoid zero count problems.

Click the github icon above or contact me to go deeper in this project!

• Python 3.6
• Anaconda
• Apache Spark 2.2.1
• Pyspark 2.2.1
• Google Cloud Platform

Implemented Random Forest classifier on malware classification via Apache Spark and processing on google cloud platform. Resulted in 98.97% classification accuracy in large dataset.

The project uses the hexadecimal binaries as documents, and classify them into one of several possible malware families. The dataset is from Microsoft Malware Classification Challenge. We have extracted seven different features from both asm and bytes files, and resulted in including unigram bytes from bytes file and segment count from asm files in our final classifier.

Click the github icon above or contact me to go deeper in this project!

• Python 3.6
• Anaconda
• tf_unet
• thunder-extraction
• CaImAn

Implemented Convolutional neural network (CNN), Non-negative matrix factorization (NMF), and Constrained non-negative matrix factorization (CNMF) to the neuron images and segmented the location of neurons. Resulted in 3.12 total scores based on the score system of neuron-finders.

The project uses the time series image datasets of neurons from CodeNeuro. Each folder of training and testing images is a single plan, and the images are numbered according to their temporal ordering. We offered three different algorithms (CNN, NMF, CNMF) based on three different python packages (Unet, thunder-extraction, CaImAn) in our github repository. Compared various methods, the resulted best result is from the CNN model using tensorflow Unet.

Click the github icon above or contact me to go deeper in this project!

• Python 3.6
• Anaconda
• Tensorflow
• OpenCV

Implemented Histogram of Gradient (HOG) descriptors, Optical Flow, Convolutional Neural Networks (CNN) to the cilium images and extracted the ciliary motion. Resulted in 20.6449 mean IoU based on the values of intersection over union of prediction and labels.

This project uses the time series grayscale 8-bit images of cilia biopsies taken with DIC optics published in the study Automated identification of abnormal respiratory ciliary motion in nasal biopsies. For each video, they are formed by 100 successive frames and performed as a 0.5 seconds real-time video. We are only intereted in cilia segmentation, the misclassification of the cells and backgrounds are note considered here.

We offered four different algorithms (HOG+SVM, Flatten images +SVM, Optical Flow, CNN) based on different directoy in our repository. (We have modulized our algorithms in the Github repository.) Compared various methods, the best result was generated from the Optical Flow in OpenCV.

Click the github icon above or contact me to go deeper in this project!

• Python 3.6
• Anaconda
• keras
• itsmr

Implemented autoregressive integrated moving average (ARIMA) model and long short-term memory (LSTM) model to web page visits time series data of Wikipedia website. Resulted in 38.8875 of mean SMAPE score of ARIMA model and 53.8052 of mean SMAPE score of LSTM model.

The project uses the web traffic time series data from the Kaggle competition Web Traffic Time Series Forecasting. Each row represents a time series of a specific page on Wikipedia from 07/31/15 to 12/31/16 for training set 1 and to 09/01/17 for training set 2. We offered two different algorithms (ARIMA and LSTM) based on two different python packages (repackaged itsmpy and keras) in our github repository and website. We have offered the PROs and CONs of each algorithm in our repo and website.

Click the github and link icon above or contact me to go deeper in this project!