Machine Learning

Machine learning workflow.

1. Asking the right questions
2. Preparing data
3. Selecting the algorithm
4. Training the model - Letting specific data teach a Machine Learning algorithm to create a specific prediction model.
5. Testing the modely

pip install pandas matplotlib scikit-learn

Predictable solution based on models built of data

1. Supervised and Unsupervised

In supervised ML, the model contain features and values to be predicted.

Example: House price predictions Value: Price Features: Size:1610, Bedrooms:3, Year Built:2007 Here a model is created using the features and trained against the price

In Unsupervised, cluster of like data is being analyzed to identify groups of data that share the same trait.

Problem

Predict if a person will get diabetes

Solution

1. Understand the features in data
2. Identify critical features
3. Focus on at risk population
4. Select data source. Pima India Diabetes study is a good source

Results

1. Binary result(True or False)
2. Genetic difference are a factor
3. 70% Accuracy is a common target

Using Pima India Diabetes data, predict with 70% or great accuracy, which people will develop diabetes

Disease prediction, Medical research practices, Unknown variations between people, likehood is used

Data Rule #3 Accurately predicting rare events is difficult

Data Rule #4 Track how you manipulate data

Selecting Algorithm

The algorithm use the result of analytic to create a model that can be trained for future predictions

1. Compare factors - Learning type(prediction model), result(continuous values, price=A*#bedroom+B*size| classification: discrete values, small, medium, large), complexity, basics vs enhanced Solution requires predictions. The outcome is binary(True/False) Hence, algorithm must support classification(binary classification). This reduces algorithm from 50 < 28 < 20
2. Complexity(Simple. Eliminate "esemble" algorithms, container, multiple child algorithms) Boost performance algorithms reduces algorithm from 20 to 14. 50 < 28 < 20 < 14
3. Basic vs Enhanced Enhanced(variation of basic,performance improvements, additional functionality, more complex) Basic(simpler) reduces algorithm to 3: 50 < 28 < 20 < 14 < 3[Naive Bayes, Logistic Regression, Decision Tree]

Algorithms

1. Naive Bayes - Based on likelihood and probability Fast - up to 100X faster Stable to data changes
2. Logistic Regression - Confusing name, binary result, weights relationship between features resulting in 1/0 output
3. Decision Tree - uses a binary tree structure to make decision. Node contains decision. Requires enough data to determine nodes and splits

Selection is based on Learning(Supervised), Result(Binary classification), Non-ensemble, Basic and finally select Naive Bayes for training

Training the Model

Datasets are always divided into 2. First 70% for training and the other 30% for testing.

Scikit-learn

Designed to work with Numpy, SciPy and Pandas. It contains toolset for training and evaluation tasks:

• Data splitting
• Pre-processing
• Feature Selection
• Model training
• Model tuning