Project

H-1B Visa Status Classification and Classifier Analysing

View Poster, View Abstract
Classifier Performance Analysis through H-1B Visa Certification Status Classification.

• Data Exploration: explore how individual features (employer, wage, worksite, etc.) related to visa certified rate using Sklearn.
  
• Feature Engineering: remove highly correlated fields, standardize numerical data and one-hot-encode categorical data.
  
• Classification: train data using Random Forest, Nearest Neighbors, SVM, MLP, Logistic Regression, and AdaBoost. Predict the results and analyze the performance of models.
  
• Integration: explore voting classifier and stacking classifier, integrate seven classifiers listed above, achieve 91%+ accuracy, 90%+ precision and 99%+ recall.

Question-Answering System Word Embedding Optimization

View Code
Optimizing the Question-Answering System by replacing the exsisting word embedding to a self-trained word embedding.

• Datasets: there are three datasets, SQuAD, NewsQA, and BioASQ. Initially, the model used GloVe embedding. I trained three embeddings associated with three datasets.
  
• Word Extraction: extracted words from the datasets and organized words in sentences, which each sentence is a list of strings.
  
• Vector Training: trained the word vectors using the Gensim Word2Vec framework, applied skip-gram algorithm with negative sampling. Stored word-vector mappings as .txt file.
  
• Similarity Comparison: selected five common words from each dataset, extracted the words that are most similar to the selected words based on each embedding. Compared the similar words generated by each model.
  
• Result: decreased negative loss by around 5 percent, decreased out-of-vocabulary rate from 10% to less than 0.1%.

Russell 1000 Future Income Growth Prediction

View Project Report
Estimated the net income growth rate of Russell 1000 companies for the next quarter.

• Feature Engineering: selected 10 of 192 features using similarity matrices to find less correlated features.
  
• Regression: grouped the data by industries and used regression models (Linear Regression, Neural Networks, SV Regression, Radius Nearest Neighbor Regression) to train the data and predict future net-income growth.
  
• Prediction: selected the model with minimum mean square error and used the result predicted by this model as the final result.
  
• Evaluation: The models decreased the MSE by 70% compared to naive prediction (output average growth rate as prediction).

Sentiment Analysis

Interpretation and classification of emotions (positive and negative) within text data using natural language processing techniques.

• Perceptron: implemented a perceptron classifier with a bag-of-words unigram and bigram feature extraction.
  
• Logistic Regression: implemented Logistic Regression with the same feature extractions and analyzed the log-likelihood.
  
• Neural Networks: implemented Feedforward Neural Network with Glove word embedding and tried to maximize negative log-likelihood and minimize losses.
  
• Used the three models for predicting sentiment example sentences, all the models achieved above 75% accuracy.

Decision Tree

• Built a decision tree for data mining using Python that predicted the result of new coming data.
  
• Calculated the entropy of each feature and obtained the best information gain, split the branches by the threshold with the best gain.
  
• Used 5-fold cross-validation to train and test the decision tree. Completed the decision tree with 65% accuracy.