Intro to Machine Learning

Introduction to Machine Learning

• Machine learning can be applied to disease prediction systems, price estimation, recommendation systems, and time series analysis
• Suppose you run an online software store and collect data for each purchase such as computer type, web browser, country and time
• A core problem involves predicting what the next person will buy based on collected data
• One can solve the prediction problem, given the data that seems to be similar, and similar people often act similarly
• Supervised learning is a type of machine learning
• Learning improves performance on a task through remembering, adapting and generalizing
• Generalizing involves transferring knowledge from one context to use in different situations
• Machine learning enables computers to modify or adapt to make more accurate predictions, actions and control robots
• Machine learning accuracy is measured by how well actions reflect correct ones
• Imagine playing Scrabble against a computer; initially, you might win every game before the computer adapts and learns to beat you
• Computer learns to beat a player and use the same strategy against other players

Definition of Machine Learning

• A computer learns from experience E relative to tasks T and performance measure P, if its performance at tasks in T improves with experience E, as measured by P
• Task: behavior or task to be improved, such as classification, acting in an environment
• Data: experiences to improve performance in the task
• Measures of improvement: increasing accuracy in prediction, acquiring new speed and efficiency

How Computer Science, Data Science, and ML relate

• Computer Science comprises machine learning
• Machine learning sits at the intersection of computer science, math, and statistics
• Data analysis, traditional software, data science, and business domain expertise combine with computer science, maths, and statistics

Applications of Machine Learning

• Machine learning is used in: medicine for disease diagnosis, computer vision, robotic control, natural language processing, speech recognition, machine translation, finance, and fraud detection

Algorithmic vs Machine Learning Solutions

• algorithmic solutions use data as input to a program to get output
• Machine learning solution uses data and output to get a program

Learner Design

• To design a learner: choose the training experience, target function, how to represent the target function, and a learning algorithm to infer it

Types of Machine Learning

• Supervised Learning
• Unsupervised Learning
• Reinforcement Learning
• Evolutionary Learning

Supervised Learning

• In supervised learning, a training set of examples with correct responses can allow generalizing correctly to possible inputs
• Supervised learning is learning from exemplars
• Supervised learning is represented as set of data, the inputs, the targets and the index suggests lots of pieces of data indexed by i running from 1 to some upper limit N

Supervised learning - Regression

• Regression assumes values come from a function to find the function, giving the output value of y for any value of x
• Regression in statistics describes a curve so it passes close to the data-points
• Regression is a problem of function approximation or interpolation, working out the value between known values

Supervised learning - Classification

• Classification problem consists of input vectors to decide corresponding classes based on training
• With classification, each example belongs to only one class when the set of possible output space is complete
• Two constraints about classification aren't necessarily realistic, since examples might belong partially to two different classes
• Classification can include classifying if an email is spam, recognize handwritten characters, and classifying user behavior to determine churn

Classification Vs Regression

• Input features X1, ..., Xn and a target feature Y uses training examples, where the value of the input features is given for each given example
• The new example values predict target features for a new example, classification of Y is discrete, and regression of Y is continuous

Unsupervised Learning

• Unsupervised learning is the type of machine learning that happens without human supervision. A machine tries to find any patterns in data by itself.
• Applicable in clustering and association problems
• Results may be less accuracte than supervised learning

Types of Unsupervised Learning

• Applicable to find clustering and associations

Supervised vs Unsupervised Learning

• With unsupervised, machine learns without human supervision
• Supervised learning is a type of machine learning that happens under human supervision. Labels input data with answer keys showing a machine the desired outputs

Reinforcement Learning

• Reinforcement learning is between supervised and unsupervised learning
• Algorithm is told when the answer is wrong but not how to correct
• Algorithm must work out how to get the answer right by trial and error
• Reinforcement learning involves a monitor that scores the answer and suggest improvements with a critic

Evolutionary Learning

• Evolutionary learning adapts biological organisms to improve survival rates
• Evolutionary learning draws inspiration from biological evolution, particularly natural selection
• Evolutionary learning also optimizes solutions to complex problems using algorithms that mimic the process of natural selection

Steps in Evolutionary Learning

• Generate candidate solutions randomly or heuristically through initialization
• Evaluate each individual in the population based on its performance in solving the problem, which is done using a fitness function
• Select individuals with higher fitness values for the next stage of evolution
• Use selected individuals to generate offspring through processes like mutation, crossover, or recombination for offspring generation
• Offspring replace some individuals in the existing population by criteria and strategies such as elitism
• Repeat steps 2-5 for multiple generations until a maximum number is reached

The Machine Learning Workflow

• The workflow includes data collection, data preparation, choosing learning algorithm, and training
• Evaluating the model determines the best predictions

The Machine Learning Process

• Collect and prepare data
• Identifying the best features for examing the problem through feature selection
• Choose algorithm along with knowledge of its use
• Parameter and Model Selection to identify appropriate values either manually or via experimentation
• Use dataset, algorithm and parameters, alongside computational resources to build a model of the data in order to predict the outputs on new data during training
• Before deployment evaluate for accuracy on data not trained on

Terminology used in ML

• Inputs are the data fed into the algorithm
• Machine learning algorithms take input values to produce an output as an answer for the input vector
• An input vector consists of real numbers
• Numbers of vector elements is dimensionality of the input

Notation

• Vector x has elements (x1, x2, ..., xm) or xi, where i runs from 1 to the number of input dimensions, m
• wij are the weighted connections between nodes i and j
• Weights are analogous to synapses in the brain, which are arranged into a matrix W
• Output vector is y, with elements yj, where j runs from 1 to the number of output dimensions, n
• y(x,W) is the set of weights of the network
• Target vector is t, with elements tj, where j runs from 1 to the number of output dimensions, n
• t are the extra data for supervised learning to 'correct' the answers the algorithm gives
• Activation function describes the firing of the neuron as a response to the weighted inputs
• Error (E) compute inaccuracies of the network as a function of outputs (y) and targets (t)
• Weight Space is thinking about the weights that connect into a particular neuron and plotting the strengths of the weights using one axis for each weight

The Curse of Dimensionality

• The essence of the curse is the realization that as the number of dimensions increases, the volume of the unit hypersphere does not increase with it
• Volume of Hypersphere is relative to its dimensions
• The unit hypersphere is the region we get if we start at the origin and draw all the points that are 1 unit away
• In 2 dimensions we get a circle, in 3 dimensions its a sphere, and then its a hypersphere for higher dimensions

Overfitting

• Training too long, can cause overfitting the data, and learning noise and inaccuracies
• Model becomes complicated and cannot generalize
• Validate the generalization aspect and learning at each time step, by stopping the learning process before algorithm overfits
• Validation set of data can be used to validate the learning

Training, Testing and Validation Sets

• Three datasets are needed: training set to train an algorithm, the validation set to keep track of how well it is doing as it learns, and the test set to produce results
• 50:25:25 is the proportion of training to testing to validation data
• Data splitting can matter, some is presented with datapoints in class 1, then class 2 and so on
• If first few points are training sets, results will be bad, since the training did not see all classes
• can reorder the data randomly first to assign each data-point randomly

Multi-fold Cross-Validation

• Dataset randomly partitioned into K subsets and 1 subset as validation and the algorithm trained on the others
• A new model is trained after a different subset is left out, and repeat the process for all subsets.
• The model with the lowest validation error is tested and used

The Confusion Matrix

• Confusion matrix can work out whether the algorithm results were good
• Confusion matrix is for classification problems and is a square matrix
• Lists possible classes in both horizontal and vertical directions that can list along the top of a table as outputs and on the left hand side as the targets
• Matrix element at (i, j) tells us how many input patterns were put into class i by the targets, but class j by the algorithm
• Algorithm performance measured and compared
• The elements of the leading diagonal relate to the data classified correctly

Accuracy Metrics - Confusion Matrix

• A true positive accurately predicts the positive class
• A true negative accurately predicts the negative class
• A false positive are inaccurately predicts positive class
• A false negative inaccurately predicts the negative class

Accuracy Metrics - Example

• In the context of a wolf example, a "wolf" is a positive class, and "no wolf" is a negative class

Accuracy Definition

• Accuracy is sum of true positives and negatives divided by total examples:
• Accuracy = (#TP + #TN) / (#TP + #FP + #TN + #FN)
• Sensitivity = #TP / (#TP + #FN)
• Specificity = #TN / (#TN + #FP)
• Precision = #TP / (#TP + #FP)
• Recall = #TP / (#TP + #FN)
• F1 = #TP / (#TP + (#FN + #FP)/2

Receiver Operator Characteristic (ROC) curve

• ROC curve plots on the y axis the percentage of true positives against false positives on x-axis
• A single classifier run produces a single point on plot, with point (0,1), 100% true positives and 0% false positives being a perfect classifier
• Any classifier lying from (0,0) to (1,1) functions as at chance level and positive and negative classes are common, wasting learning effort

Unbalanced Datasets

• In a balanced dataset assume there are similar numbers of positive and negative examples
• Balanced accuracy computed with:
• MCC = (#TP x #TN - #FP x #FN) / sqrt((#TP + #FP)(#TP + #FN)(#TN + #FP)(#TN + #FN))

Measurement Precision

• Measurement precision evaluates accuracy of a learning system using word "precision" for measurement systems
• Measurement happens by feeding in inputs and look at the outputs
• Compare to target value, can measure if fed in a set of similar inputs, so expect similar output
• Algorithms variability is the precision, that tells how repeatable its predictions are

Averages

• Mean is the average of a dataset by adding the points and dividing by number of points
• Median can be found by middle value after sorting the dataset
• Mode can be found by the most common value and counting times an element appears

Variance and Standard Deviation

• Variance of numbers in a set is measured by how spread out values are
• Variance computed with sum of squared distances between each element and expected values of the set (mean, μ)
• Square root of variance is standard deviation, represented as the sigma symbol

Covariance

• Covariance is used to look at how two variables vary together
• Covariance is a measure computed by cov({xi}, {yi}) = E({xi} – μ)E({yi} – ν)
• If two variables are independent, the covariance is 0 and known to be uncorrelated
• If both variables both increase and decrease at the same time, the covariance is positive
• If variables display opposite patterns, covariance is negative

Calculate the average value for each variable

• μ = (1,245 + 1,415 + 1,312 + 1,427 + 1,510 + 1,590) / 6 => μ = 1,416.5
• v = (100 + 123 + 129 + 143 + 150 + 197) / 6 => v = 140.3

Steps in finding differences

• Find the difference between each value and the mean for both variables

Covariance between all pairs of variables in a Dataset.

• Covariance applied to used to look at the correlation between all pairs of variables within a set of data
• To compute, it needs to compute each then put together in the convariance matrix
• Element xi is a column vector describing the elements of the ith variable and μi is its mean
• Matrix is square, elements from leading elements equal to variances, equal to covariance

Mahalanobis Distance

• Distances between the center of the data indicate whether something would belong to a dataset
• This assesses relationship to the spread of the data-points
• Mahalanobis constructs a distance measure taking account, called the 1936 is written is:
• DM(x) = sqrt((x – μ)TΣ-1(x – μ))

The Gaussian

• Gaussian distribution is key because of the Central Limit Theorem which will add up to something Gaussian

The Bias-Variance Tradeoff

• A model can be bad for 2 reasons: not accurate, not precise/has variation in results
• First of the reasons is the bias and the second is the variance
• With more complex classifiers improving the bias results in higher variance
• Making a model more specific result in reducing variance, increasing the bias