https://github.com/cdvel/ml-tutorials

Machine Learning Tutorials in Python (Notes and Implementations)
https://github.com/cdvel/ml-tutorials

Last synced: 8 months ago
JSON representation

Machine Learning Tutorials in Python (Notes and Implementations)

• Host: GitHub
• URL: https://github.com/cdvel/ml-tutorials
• Owner: cdvel
• Created: 2015-03-06T09:18:27.000Z (about 11 years ago)
• Default Branch: master
• Last Pushed: 2015-03-25T09:49:19.000Z (about 11 years ago)
• Last Synced: 2025-06-04T05:43:48.960Z (12 months ago)
• Language: Python
• Homepage:
• Size: 172 KB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# ml-tutorials

1. [k-Nearest neighbors](http://machinelearningmastery.com/tutorial-to-implement-k-nearest-neighbors-in-python-from-scratch/)

  - Finds the N most similar elements based on euclidean, hammington  etc distance

  * Instance-based

    - model the problem using data instances(rows) in order to make predictive decisions. 

    - In kNN all observations are retained as part of our model = extreme instance-based

  * Competitive learning 

    - internally model elements(instance) compete in order to make a predictive decision.

    - objective similarity measure between instances causes each instance to compete to be most similar to a given unseen data instance and contribute

  * Lazy learning

    - The algorithm doesn't build a model until the time of prediction is required

    - Only relevant data to the unseen data (localized model)

    - Can be computationally expensive to repeat over larger training sets

  - Makes no assumption over the data, only that a distance measure can be calculated. Non-parametric or non-linear, doesn't assume functional form.

2. [Naive bayes] (http://machinelearningmastery.com/naive-bayes-classifier-scratch-python/)  

  - Suits classification problems. Uses probabilities of each attribute belonging to each class to make a prediction

  - Fast and effective supervised learning for probabilistic prediction

  * The Model 

    - A summary of data in the training set

    1. mean

    2. standard deviation for each (no. attributes * class values) 

    3. calculates probability of a specific attribute belonging to each class value      

  * Assumptions

    - Independent probabilities between attributes of a given classs

    - Numerical attributes are normally distributed 

  * Conditional probability

    - Probability of a class given a value of an attribute

    - The product of _conditional probabilities_ for each attribute = probability of an instance belonging to a class

  * Prediction

    - Calculate probabilities of instance belonging to class

    - Pick class with highest probability

  * Implementations

    - Encoding using [Log Probabilities](https://en.wikipedia.org/wiki/Log_probability):

      - Reduces risk of floating point underflow (values too small to be represented)

      - More efficient by using summation of log probabilities instead of product of probabilities

      - Observations with the Iris dataset: 

        - Less accurate 

        - Faster

        - Log = (0.5s - 0.8s, accuracy=66%) vs Prob = (0.6 - 0.8, accuracy=74%)

        - Numerical stability (accuracy is more consistent)

    - using categorical data (ratios)

    - numeric attributes (with normal distribution)

  - [How to get best from Naive bayes] (http://machinelearningmastery.com/better-naive-bayes/)