(Date: Sep. 30, 2019)

The Iris data has 150 samples, 4 attributes, and 3 species (target) and they are confusing to distinguish with the naked eye. The goal is to predict the class of the flower as accurately as possible and Linear Discriminant Analysis(LDA) model, achieved 97.4% of accuracy on test set and only suffered 0.5% variance from training set.

(Date: Oct. 5, 2019)

This project is to answer "should this loan be approved or denied?". From a company's perspective, you want to minimize the risk of not getting money back and of losing possible customers. The data had only 614 examples and the model was only trained on 591 examples. The data consisted of more categorical features than numerical thus through data transformation, data became sparse and it turned out data points were jumbled together regardless of its class and therefore it is hard to find the decision boundary. Extreme Gradient Boost (XGB) and SVM models achieved 0.83 of mean average precision and 0.85 of accuracy equally. With confusion matrix, the model clearly did excellent job in filtering out False Negatives reducing the risk of losing customers, but did relatively poorly in filtering out False Positives.

(Date: Oct. 17, 2019)

This project aims at providing "smart pricing" for New York City Airbnbs. The data is composed of 48,895 rows and 16 columns. The optimal model achieves 95% coverage with $36.07 MAE (Mean Absolute Error). The reason the model could not achieve 100% coverage originates from the fact that high pricing airbnbs are rare and therefore not many training models provided for the model to be trained on. For predicted values above 95 percentile training price $350 should be examined by human experts.

(Date: Oct. 23, 2019)

Predictions of sunspot number are also very important in planning space-related activities particularly for low Earth orbiting spacecraft. This project predicts monthly sunspot numbers. The data is composed of the observations over 18 years (2001 - 2018) from Australia. The CRNN (Convolutional Recurrent Neural Network) predicts the number of sunspot with 1.78 MAE (mean absolute error).

(Date: Nov. 7, 2019)

This project aims to develop a model that an predict the market price for Google. The data is composed of 13 years of stock data (2006-01-01 to 2018-01-01). With multi-layer LSTM model, I achieved 43.67 MAE. The model underperformed on the test set (training set: 17.17 MAE), but it is due to the configuration of the data. The data shows very mild upward trend until 2012 and then exponential-like trend after 2012, which means, we do not have enough data to read this trend more precisely. With more accumulation of data throughout the years, the performance level will be improved.

(Date: Feb. 13, 2020)

All 52 cards lie face-down; half the cards are red and half are black. We proceed to flip the cards over one by one. If the last card you flip over is red, you will win a dollar and otherwise lose a dollar. You can halt the game at any time and flip the last card which will serve as the final card. We can play the game as many times as you like. The deck will be reshuffled every time. After each round, we exchange money. The project is aiming to figure out the best probabilistic approach to winning the game with the help of simulation. After experiments I came to conclusion that "No matter what we do, our likelihood of winning remains 50-50. Therefore, the most optimal strategy we can off is to pick the first card in the shuffled deck."

(Date: March. 19, 2020)

The owner of a business wants to know if there exists a color among 30 different colors that draws significantly more ad-clicks than blue which is current color of text. With naive Permutation test 'Ultramarine' came out as significant but after Bonferonni correction taking in to account comparisons between blue and 29 other colors, the p-value was not statistically significant. If the owner had simply compared blue to the other 5, say, colors, then maybe we would have obtained a statistically significant result.