I am JasonA Data Scientist

Using SQL to analyze website performance is a daily routine for data scientists. It's not necessary to always use machine learning and deep learning models to create predictive models. Solving business problems is the top task for us rather than creating models. In this project, I use an e-commerce website dataset to analyze some website problems such as traffic source, bounce rate and so on. Even though I'd like to create some predictive models in real life, I only use SQL in this project.


Software used: SQL
Click the "Details" button for more details.

This is a real fintech company that wants to provide its customers with a paid mobile app subscription that will allow them to track all of their finances in one place. To attract customers, the company releases a free version of their app with some main features unlocked. The purpose of this project is to identify which users will most likely not enroll in the paid product so that additional offers can be given to them. I used several models and finally got 100% accuracy. This is just an example of how I find out the company's potential customers. The techniques I used here could be applied in other companies as well.
In real life, things are more complicated, so I also have some last thoughts in the end.


Software used: Python

Packages used: Sklearn, Tensorflow, Keras, Seaborn

Click the "Details" button for more details.

Time series forecasting is a crucial topic for data scientists. Autoregressive integrated moving average (ARIMA) is one of the famous linear statistical models for time series forecasting. It usually outperforms machine learning and deep learning models for one-step forecasting on univariate datasets. There are different variations of ARIMA. In this project, I use one type of ARIMA model which is SARIMAX, Seasonal Auto-Regressive Integrated Moving Average with eXogenous factors. A restaurant visitors dataset has been used in this project.

Software used: Python
Packages used: statsmodel, pmdarima
Click the "Details" button for more details.

Data scientists deal with big data by using Spark which is very powerful and user-friendly. Spark 2.0 has shifted towards a dataframe syntax which is just like pandas and Spark's MLlib library is similar to sklearn library. You also can write SQL queries by using Spark. Since I've had many machine learning and SQL projects, I used Spark MLlib to create a spam filter in this Natural Language Processing project.

Software used: VirtualBox, Ubuntu, Jupyter Notebook
Packages used: Pyspark, Spark MLlib
Click the "Details" button for more details.

A/B testing is a very popular topic among companies. As a data scientist, there are several things that I need to check before I conduct A/B testing. The first half of this project is the workflow I'd go over every time. There's also an example of A/B testing in the second half.

If an A/B testing experiment has multiple treatment groups, the cost of testing each treatment group and the probability of false positives will increase. When there are multiple treatments groups, I usually use two types of reinforcement learning, Upper Confidence Bound (UCB) Algorithm and Thompson Sampling to quickly find out the best treatment group by using the minimum amount of time. You can see how I use reinforcement learning to conduct A/B testing in A/B Testing Reinforcement Learning Project



Software used: Tableau

Click the "Details" button for more details.

Self-Organizing Map is a type of unsupervised deep learning model. The purpose of using SOMs is to reduce the dimension of our dataset and create visuals. In this project, I combined SOMs and the Artificial Neural Network to create a model for fraud detection. We don't want to see that our customers are cheated, and then we perform supervised learning to protect them in the future. Fraud has to be stopped before it happens.


Software used: Python, Pycharm
Packages used: Tensorflow, Keras, Sklearn, MiniSom, pylab,
Click the "Details" button for more details.

This is a preprocessed dataset. There are 28 features not interpretable because the personal information of our customers has been transformed and kept anonymous. The "Class" feature is whether customers have been defrauded or not, 0 = no, 1 = yes. I used undersampling and oversampling in this case because this dataset contains more than 280,000 observations, but only about 400 people have been defrauded, which just accounts for 0.14% of the dataset.


Software used: Python

Packages used: Sklearn, Seaborn, Tensorflow, Keras, imblearn

Click the "Details" button for more details.

I used AWS Redshift to investigate a drop in user engagement for Yammer. Yammer is an enterprise social networking service used for private communication within organizations. This is a relatively small dataset, with more than 34,000 rows of records and 5 tables, but I used this demo to show you my thinking process when I face a business problem.



Software used: AWS Redshift, AWS S3, AWS IAM

Click the "Details" button for more details.