Machine learning

Everything until deep learning

Probability

Study of uncertainty and randomness, used to model and analyze uncertainty in data.

A form of regularization

Ridge regression

Rows on a confusion matrix

Correspond to what is predicted

Collumns on a confusion matrix

Correspond to the known truth

The sensitivity Metric equation

True positives divided by the sum of true positives and false negatives

The Specificity metric equation

True negatives divided by true negatives plus false positives

if sensitivity = 0,81 what does it mean

example: tells us that 81% of the people with heart disease were correctly identifies by the logistic regression model

If specificity = 0.85 what does it mean

It means that 85% of the people without heart disease were correctly identified

When a correlation matrix has more than 2 rows, how do we calculate the sensitivity

We sum the false negatives

What is the function of specificity and sensitivity:

It helps us to decide which machine learning method would be best for our data

Sensitivity

If correcty identifying positives is the most important thing to do, which one should i choose? Sensitivity or Specificity?

If correctly identifying negatives is the most important thing, which one should I choose? Sensitivity or specificity?

Specificity

ROC

Receiver operator Characteristic

Roc funtion

To provide a simple way to summarize all the information, instead of making several confusion matrix

The y axis, in ROC, is the same thing as

Sensitivity

The x axis, in ROC, is the same thing as

Specificity

True positive rate =

Sensitivity

False positive rate =

Specificity

In another words, ROC allows us to

Set the right threshold

When specificity and sensitivity are equal,

the diagonal line shows where True positive rate = False positive rate

The ROC summarizes…

All of the confusion matrices that each threshold produced

AUC

Area under the curve

AUC function

To compare one ROC curve to another

Precision equation

True positives / true positives + false positives

Precision

the proportion of positive results that were correctly classified

Precision is not affected by imbalance because

It does not include the number of true negatives

Example when imbalance occurs

When studying a rare disease. In this case, the study will contain many more people without the disease than with the disease

ROC Curves make it easy to

Identify the best threshold for making a decision

AUC curves make it easy to

to decide which categorization method is better

Entropy can also be used to

Build classification trees

Entropy is also the basis of

Mutual Information

Mutual Information

Quantifies the relationship between 2 things

Entropy is also the basis of

Relative entropy ( the kullback leibler distance) and Cross entropy

Entropy is used to

quantify similarities and differences

If the probability is low, the surprise is

high

If the probability is high, the surprise is

low

The entropy of the result of X is

The expected surprise everytime we try the data

Entropy IS

The expected value of the surprise

We can rewrite entropy using

The sigma notation

Equation for surprise

Equation for entropy

Entropy

Is the log for the inverse of the probability

R2 *R Squared does not work for

Binary data, yes or no

R squared works for

Continuous data

Mutual information is

A numeric value that gives us a sense of how closely related two variables are

Equation for mutual information

Joint probabilities

The probability of two things occuring at the same time

Marginal Probabiities

The opposite of joint probability, is the probability of one thing occuring

Least sqaures =

Linear regression

squaring ensures

That each term is positive

Sum of Squared Residuals

How well the line fits the data

Sum of Squared Residuals function

The residuals are the differences between the real data and the line, and we are summing the square of these values

The Sum of square residuals must be

as low as possible

First step when working with bias and variance

Split the data in 2 sets, one for training and one for testing

How do we find the optimal rotation for the line

We take the derivative of the function. The derivative tells us the slope of the function at every point

Least squares final line

Result of the final line, that minimizes the distance between it and the real data

The first thing you do in linear regression

Use least squares to fit a line to the data

The second thing you do in linear regression

calculate r squared

The third thing you do in linear regression

calculate a p value for R

Residual

The distance from the line to a data point

SS(Mean)

Sum of squares around the mean

SS(Fit)

Sum of squares around the least squares fit

Linear regression is also called:

Least squares

What is Bias

Inability for a machine learning method like linear regression to capture the true relationship

How do we calculate how the lines will fit the training set:

By calculating the sum of squares. We measure how far the dots are from the main line

How do we calculate how the lines will fit the testing set:

Overfit

When the line at the training set data fits well, but not it does not fit well on the testing set

Ideal algorithm

Low bias, accurate on the true relationship

Low variability

Producing consistent predictions across different datasets

Result of least squares determination value for the equation parameters

it minimizes The sum of the square residuals

Y= Y-intercept + slope X

Linear regression

Y = Y-intercept + slope x + slope z

Multiple regression

Equation for R2 r squared

R2 = ss(mean) - ss(fit)

---

ss(mean)

Goal of a t test

Compare means and see if they are significantly different from each other

Odds are NOT

Probabilities

ODDS are

the ration of something happening ex. the team winning

---

to something not happening, ex. the team NOT winning

Logit function

Log of the ration of the probabilities and formas the basis for logistic regression

log(odds)

Log of the odds

log odds use?

Log odds is useful to determine probabilitirs about win/lose, yes/no, or true/false

Odds ratio

ex>

Relationship between odds ration and the log(odds ratio)

They indicate a relationship between 2 things, ex a relationship between the mutated gene and cancer, like weather or not having a mutated gene increases the odds of having cancer

Tests used to determine p values for log (odds ratio)

Fisher`s exact test, chi square test and the wald test

Large r squared implies…

A large effect

Machine Learning

Using data to predict something

Example of continous data

Weight and age

Example of discrete data

Genotype and astrological sign

Which curve is better? the one with maximum likelihood or minimum?

Maximum likelihood

Type of regression used to asses what variables are useful for classifying samples

Logistic regression

Components of GLM - Generalized Linear Models

Logistic regression and Linear models

The slope indicates

the rate at which the probability of a particular event occurring changes as the independent variable changes.

Logit function

Log(p)

---

1-p p is the middle line

If the coefficient estimate in logistic regression is negative, the odds are

against, Ex if you don't weigh anything, the odds are against you being obese

if the coefficient estimate is positive, that means that

For every unit of x gaines, the odds of y increases by number on the coefficient

In logistic regression, by using the z value, how do we confirm that it is statistically significant?

Greater than 2. ex. 2.255 with a p-value less than 0.05 ex 0.0241

What the difference between the coeeficitents used for linear models and logistic regression?

Is the exact same, except the coefficients are in terms of log odds

In logistic regression, what is the scale of the coefficients?

Log(odds)

How lines are fit in Linear regression?

by using least squares, measuring the residuals, the distances between the data and the line, and then squared them so that the negative value do not cancel out positive values

Line with the smallest sum of squared residuals is

The best line

Line with the biggest sum of squared residuals is

The worst line