haowang47 / bios-823-2020

Syllabus for BIOS-823-2020

BIOS 823 is designed for people who want to work in a data science team, especially in a healthcare setting. Hence the course focuses on the skills and knowledge needed to contribute to the main roles in a data science team:

• Making data available and accessible for analysis
• Processing data and performing exploratory data analysis
• Statistical inference and machine learning
• Scaling up to big data and deploying automated workflows

You will learn how to use the essential tools for each of these roles effectively. The emphasis is on conceptual understanding, and we will highlight only selected software packages that exemplify each role. Once you understand the concepts, it should be simple to self-learn other similar packages, and you have the opportunity to do so in projects. Applications of data science to the analysis of structured, semi-structured and unstructured data, especially from biomedical contexts, will be interleaved into the course.

Prerequisites

• You have a background in calculus and linear algebra
• You have a background in basic statistics inference
• You have a background in basic machine learning
• You are fluent in Python
• You are familiar with the use of git for version control
• You enjoy coding and playing with data

Course repository is at https://github.com/cliburn/bios-823-2020

Due to the wide-ranging nature of the syllabus, there is no course textbook. However, recommendations for useful books for each major topic will be suggested for those who want to explore the topic in more depth.

Learning objectives

• How to contribute as a data science team member

• How to process data and perform exploratory data analysis

• How to manage data with SQL and NoSQL databases

• How to build machine learning pipelines

• How to scale and deploy data science products

Grading

• 50% homework
• 25% exams
• 25% final project

For letter grade

• A: 90 - 100
• B: 75 - 89
• C: 60 - 74
• F: 59 and below

Office hours

• Michael Gao (TA): Friday 2:30-3:30 PM
• Cliburn Chan (Instructor): Friday 4:00-5:00 PM
  • Please email me if you plan to attend my office hour so I remeber to log into Zoom!

Data Science

As with R, the basic foundations of data science is the creation of tidy data formats that can then be visualized using a grammar of graphics. We will look at the use of the foundational pandas package for data processing and exploratory data visualization, then at common formats for data sharing, and finally at REST and GraphQL application programming interfaces (API) for transfer of structured data resources.

Data processing and EDA

Typical packages: pandas, plotnine, plotly, streamlit

• Warm-up exercises with python and numpy
• Data manipulation with pandas
• Primer of data visualization
• Exploratory data analysis with plotnine and plotly
• Dashboards with streamlit

Portable data formats

Typical packages: json, ElementTree, h5py, pyarrow, parquet

• Text
• Tabular data
• Structured data
• REST APIs

SQL databases

Most data in healthcare settings still reside in relational database's, especially in enterprise data warehouses that have been created to serve as portals to electronic health records. Hence, an essential skill for a data scientist is the use of SQL, especially for constructing data queries. We will also examine examples of a few NoSQL databases, and review their comparative advantages and disadvantages relative to relational databases.

Typical packages: sqlite, sqlalchemy

• Relational databases are based on set theory
• Tables, rows, columns, keys, relations, constraints
• Database schema and normalization
• CRUD
• SQL queries
  • Single table queries
  • Window function
  • Joins
  • User defined functions

NoSQL databases

Typical packages:redis, mongodb, neo4j

• Key-value
  • Uses
  • Collections
• Document
  • Uses
  • Queries
• Graph
  • Graph concepts with networkx
  • Uses
  • The Cypher language

Machine Learning

Machine learning is an essential skill for any data scientist, so we will review practical issues in classical and deep learning, including feature engineering, imbalanced data, hyper-parameter tuning, and model expandability.

Understanding ML

Typical packages: scikit-learn, imbalanced-learn, pycaret, tune-sklearn, yellowbrick, auto-sklearn

• Unsupervised learning
  • Dimension reduction
  • Clustering
  • Recommender systems
• Supervised learning
  • Basic framework for training and evaluation
  • Model families
  • Preprocessing and feature engineering
  • Hyperparameter tuning
  • Model evaluation
  • Explaining models

Modern ML frameworks

Typical packages: tensorflow2, keras, tune, autokeras, dalex

• Neural network concepts
• Working with data sets
• Effective training
• Explainable AI (XAI)

Data Engineering

Finally, we consider three aspects of data engineering - construction of pipelines using functional approaches, scaling of algorithms for performance, and the automation of workflows.

Functional programming

Typical packages: itertools, functional, toolz

• Recursion
• Lazy evaluation and iterators
• Comprehensions and generators
• Anonymous functions
• Partial application and currying
• Higher order functions: map, filter, reduce, decorators
• Pipes and chaining operations
• A digression: coconut

Distributed computing

Typical packages: spark, dask, vaex

• Concurrent, parallel, distributed
• Simple multi-core processing
• Spark concepts
• Spark SQL
• Spark ML
• Spark Streaming
• GraphFrames
• Alternatives to Spark - dask and vaex

Workflow management

Typical packages: py.test, Singularity, airflow

• Literate programming
• Source code version control
• Testing code
• Continuous integration
• Using containers
• Automating complex workflows
• Creating APIs for analysis with fastapi
• Life in the cloud