Career Profile
Self-motivated data scientist and machine learning engineer with extensive project experience in deep learning, statistical analysis, modeling, full stack web development and database design.
Experiences
Data scientist internship at Uber for Business team.
Software engineer internship at Microsoft Satori knowledge graph team.
Graduate student in Quantitative Biomedical Research Center (QBRC). My major research involved:
Projects
Machine learning and deep learning
Deep learning model to forecast customer life time value
Developed a deep learning model to forecast customer life time value from historical time series time.
Used CNN/LSTM to extract time series features.
Combined time series features with time independent (geographical, business) features.
Model extendable to forecast different time horizons (1 month, 3 months, 12 months).
Deep Learning, Convolutional neural network, Long short-term memory, Life time value, Time series modeling
Convolutional neural network model for entity relation extraction
Developed a convolutional neural network model to extract entity relations from text corpus.
Used word embedding to transform tokens into numeric vectors, and applied convolutional kernels to extract semantic features to do prediction.
Deep Learning, Convolutional neural network, Word embedding, Relation extraction, Information retrieval
DrivenData competition: predicting water pump functionality using machine learning
Developed a random forest classification model to predict water pump functionality in Africa.
This is a open data challenge hosted by DrivenData.
Machine Learning, Data challenge, Random forest, XGBoost, Classification
Deep learning model to predict anticancer drug response
Developed a deep learning model to predict anticancer drug response in lung cancer.
Used convolutional neural network framework to extract features from tens of thousands of genomic mutation locus.
Combined gene expression profile and drug information to predict drug sensitivity.
Convolutional neural network, Deep learning, Drug response
Statistical methodology and modeling
Bayesian hierarchical model to predict mRNA methylation sites
Developed a Bayesian hierarchical model to analyze methylated RNA immunoprecipitation sequencing (MeRIPseq) data.
Used zero-inflated negative binomial distribution to model sequencing read count. Modeled spatial dependency by hidden markov model.
Bayesian hierarchical model, Hidden markov model, Count data, Negative binomial distribution, Zero inflation
EM algorithm to identify genes that have tri-modal distribution
Developed an expectation maximization (EM) algorithm to deconvolute gene expression profile into three Gaussian components
(lower than normal, close to normal and higher than normal). Correlated breast cancer patients' survival with this stratification
to search for genes whose both high and lower expression are associated with worse survival outcome, which may potential have
dual role (both oncogene and tumor suppressor) in cancer development.
Expectation maximization, Gaussian mixture, Survival analysis, Breast Cancer
Bioinformatics software and web portal
Large scale profiling of RBP-circRNA interaction using CLIP-Seq data
Developed a bioinformatics software named Clirc to identify RBP-bound circRNAs through analysis of CLIP-Seq data.
Circ-RNA, CLIP-Seq, RNA-binding protein (RBP)
Ensemble-based method to infer gene network structure
Developed an ensemble method to aggregate network constructed by multiple statistical methods (correlation-, mutual information-,
Bayesian-, likelihood-based methods); Developed a web server for online network inference and visualization.
Ensemble method, Hub node, Precision matrix, Network visualization, Online tool
DIGREM: an integrated web-based platform for detecting effective multi-drug combinations
Developed a computational algorithm to predict drug combination synergistic effect with transcriptomic profile,
drug dose response curve and gene regulatory network. Developed a user friendly web server to enable online prediction.
Implement this algorithm with an R package.
Drug-Induced Genomic Response models (DIGREM), Precision matrix, Drug synergy, Online tool
Convolutional neural network model for entity relation extraction
In recent years, various large-scale knowledge bases (KBs) such as Freebase, Google’s Knowledge Graph and Microsoft’s Satori
have been built and widely used to power web search engine and question answering. These KBs mostly compose of relational facts
of entities. Although existing KBs already collected facts from various large structured database, there is still a massive
amount of real-world facts that stores in unstructured text. To enrich KBs, many efforts have been invested in automatically
extracting those unknown relational facts.
The task of relation classification (RE) can be defined as follow: given a sentence S with annotated pairs of nominals e1 and e2,
we aim to identify the relations between e1 and e2, therefore can be considered as multi-class classification problems. We applied
a convolutional neural network framework to automatically extract the lexical and sentence level features without complicated
preprocessing. Then the features are fed into softmax classifier to predict the relationship of two marked nouns.