Category Archives: Machine Learning

Starting with PySpark – configuration

Machine Learning No Comments

PySpark is a pain to configure.

For this guide I am using macOS Mojave.
Spark version 2.4.0
Python 3

Start by downloading the Spark https://spark.apache.org/downloads.html. Extract wherever – can be your home directory.

Install Java SDK. Important – some later versions don’t seem to be compatible with spark 2.4.0. Version 8 seems to work- https://www.oracle.com/technetwork/java/javase/downloads/jdk8-downloads-2133151.html

Install pyspark: pip install pyspark

Configure your zshrc/bash_profile – depending on what shell you use:

export SPARK_PATH=~/spark-2.4.0-bin-hadoop2.7
export PYSPARK_DRIVER_PYTHON="jupyter"
export PYSPARK_DRIVER_PYTHON_OPTS="notebook"

export PYSPARK_PYTHON=python3
alias snotebook='$SPARK_PATH/bin/pyspark --master local[2]'

export SPARK_HOME=~/spark-2.4.0-bin-hadoop2.7
export PATH=$SPARK_HOME/bin:$PATH
export PYTHONPATH=$SPARK_HOME/python:$SPARK_HOME/python/lib/py4j-0.10.7-src.zip:$PYTHONPATH

export PYTHONPATH=$SPARK_HOME/python/lib/py4j-0.8.2.1-src.zip:$PYTHONPATH
export PYSPARK_SUBMIT_ARGS="--master local[2] pyspark-shell"

export JAVA_HOME=$(/usr/libexec/java_home)

Remember to reload your console.

Now, when you enter pyspark on your console, it’ll open a notebook.

You can validate if Spark context is available by entering this in your new notebook:

from pyspark import SparkContext
sc = SparkContext.getOrCreate()

References: https://medium.com/@yajieli/installing-spark-pyspark-on-mac-and-fix-of-some-common-errors-355a9050f735

My Thoughts about CSE 6250 Big Data Analytics in Healthcare (taking in Spring 2019)

Machine Learning 5 Comments

This is my first class with Georgia Tech OMSCS program. I would prefer to take a different class as my first class (Machine Learning) so I have a better understanding of machine learning algorithms before trying to apply them, but as a newcomer you’re the last in a priority list. So, the other classes were completely full by the time I could make my selection.

Prerequisites

It would help if you are familiar with Python and at least some machine learning algorithms.

The second homework involves some math that requires you to use a chained derivative rule. However, the majority of the tasks are more practical.

Effort

Very intense. You’ll have to use multiple languages and tools to accomplish your homework. For this year (Spring 2019) this includes Python, a variation of SQL, Scala; Hadoop, Pig, Spark… This class is taking more of my time than what I wanted to spend on it with a family and a full-time job.

Grading

The automated code grader has bugs.

My first homework had some points taken off because of the tasks we were not even supposed to do. I contacted the teaching assistant (TA) and had a full credit restored.

The second homework had some points taken off, because they split the script into parts and ran each part separately, while my code was expecting that the whole script would run as a whole. The assignment did not mention anything about this. I again had the full credit restored after talking to teaching assistants and demonstrating that the issue was with this unstated requirement. The teaching assistant was very responsive.

For my third homework (Spark + Scala), I initially received 0 points, because I was trying out some plugins and modified the scala project file. Then I forgot to remove it, and my homework could not be run with the automated grader. This time the first TA never responded (I waited for about 4 days and followed up once), but the second TA replied right away. He manually reran my code and I only lost a few points due to the bad project file.

The last, fifth homework (PyTorch + deep learning) requires a lot of time. You can take a part in Kaggle competition with other classmates as a part of this homework. I totally sucked at this one. I think I had some bugs in the data preprocessing stage, even though I passed the included unit tests.

A note about the homework submission process – if you miss a file or make a typo, you won’t know about it until you homework is officially graded. There is no immediate feedback on submission.

Tools

Docker – there are several ways you can run your homework assignments. If you don’t want to set up your home environment for each task. I used the provided docker image (there is also an option to use Azure virtual machine, but I did not use that option).

TEX editor – I used TeXstudio on Mac. You can use a regular Word and save to pdf for homework assignments which require a written answer. But, some of them require you to type formulas. And, although I found using TEX format extremely frustrating, at least the original homework assignment is provided both in tex and pdf formats. So you can start with that provided tex file and adjust fill out the answers.

Overleaf – something I discovered at the end of the class. This is a an online LaTex editor that allows you to collaborate with other students. As long as you sign up with your Georgia Tech email, it’s free.

Professor Involvement

Nonexistent. Your only chance to see a professor is through Udacity. The professor did not answer a single question on Piazza; it was 100% TAs.

Overall Impression

There is no need to jam so many technologies in a single class. Sometimes, I felt like I was just going through different sections of the homework filling out the missing parts (they usually provide a method signature and you’re supposed to write the code), without actually understanding the bigger picture. Not a bad class, especially if you can dedicate enough time to it, but would not recommend it as your first class.

Machine Learning Algorithms Problem Types

Data Science Machine Learning No Comments

Types of problems we can solve with machine learning:

  • Regression- helps establish a relationship between one or more sets of data

    • Algorithms
      • Simple linear regression
      • Multiple Linear Regression
      • Polynomial Regression
      • Support Vector Machines (SVR)
      • Decision Tree
      • Random Forest Regression
    • Sample problem: calculate the time I get to work based on the route I take and the day of the week
  • Classification – helps us answer a yes/no type of question based on one or more sets of data

    • Algorithms
      • K Nearest Neighbors (KNN)
      • Kernel SVM
      • Logistic Regression
      • Naïve Bayes
      • Decision Tree
      • Random Forest Classification
    • Sample problem: will I be late or on time based on the route I take and the day of the week
  • Clustering – helps us discover clusters of data

    • Algorithms
      • Hierarchical Clustering
      • K Means
    • Sample problem: classify the customers into specific groups based on their income and spending
  • Association – helps determine an association among multiple events

    • Algorithms
      • Apriori
      • Eclat
    • Sample problem: if I like movie A, what other movies will likely to enjoy
  • Reinforcement – helps to better exploit while exploring

    • Algorithms
      • Thomson Sampling
      • UCB
    • Sample problem: we want to determine the most effective treatment. Instead of conduction a long-term random trial, use UCB or Thompson Sampling to determine the best treatment in a shorter interval
  • Natural Language Processing

    • Algorithms
      • Any classification algorithm, but most popular are Naïve Bayes and Random Forest
    • Sample problem: determine if an amazon review is positive or negative
  • Deep Learning – can help determine hard to establish non-linear relationships between multiple input parameters and some expected outcome

    • Algorithms
      • Artificial Neural Networks (ANN)
      • Convolutional Neural Networks (CNN) – especially helpful when processing images
    • Sample problem: based on the credit score, age, balance, salary, tenure… determine if a customer is likely to continue using your service or leave