• No results found

2.4 Machine Learning

2.4.1 Machine Learning Classification

Learn to React Data-Driven

· Unlabelled data

· No direct feedback

· Find hidden pattern

Task-Driven · Labelled data · Feedback loop · Predict outcome or future

Machine Learning

Supervised Learning Unsupervised

Learning Reinforcement Learning

· Decision process

· Reward based system

· Try learning series of actions

· Linear Regression

· Polynomial Regression

· Support Vector Regression

· Decision Trees

· Neural Network

· Support Vector Machine

· Logistic Regression

· Naïve Byes

· Linear Discriminant Analysis

· K Nearest Neighbour Regression Classification · K-mean · Fuzzy C-mean · Gaussian Mixture · Neural Network -Deep Belief -Self Organising Map

· Generative Adversarial Network

· Hidden Markov Model

· Hierarchical Clustering · Principal Component Analysis Clustering · Monte Carlo · Temporal Difference Methods · Q-Learning · SARSA algorithm Data+Task-Driven

· Few labelled + large unlabelled data

· Assumption based underlying structure

· Infer correct labels and predict outcome

Semisupervised Learning

· Generative Models

· Graph Based Models

· Heuristic Approach

· Low Density Separation

Figure 2.4: Classifications of machine learning with specified characteristics (purple box) of each category along with representative examples of algorithms.

2.4.1 Machine Learning Classification

Machine Learning can be divided into 4 categories, 1) Supervised learning [69] 2) Unsupervised learning [70] 3) Semi-supervised learning [67] [68] 4) Reinforcement learning [71] as shown in Figure 2.4. This figure explains high-level characteristics of each of the category along with some examples of machine learning methods.

1. Supervised Learning

Supervised learning uses known data sets (labelled data) as training data for making predictions. The training data includes input data as well as response values. Explicitly, a supervised learning algorithm builds a model based on the training data, which predicts future response values for new data set, such as weather forecast to predict temperature or precipitation value.

Supervised learning can be divided into two categories: Classification and Regression as shown in Figure 2.4. Classification is a way of learning from categorical response values, where data can be separated into specifically labelled classes. In other words, it can be classified as a discrete- valued output method, for example, to detect if the cancer is benign or not. By contrast,

Regression algorithm is used for predicting continuous output response, for instance, predict a

house price, and temperature prediction for the weather forecast. Some of the learning methods underClassification and Regression categories are listed in Figure 2.4.

A workflow model of supervised learning is explained in Figure 2.5. A process of model generation starts with feature extraction from available labelled data. Extracted features are then used to fit a model which is then trained on a training data set in an iterative fashion to update the model. To check the accuracy of data, the model is tested on test data set and if satisfied then used for future prediction.

Raw Data (Labelled)

Feature

Extraction Fit a Model

Train a Model

Test a Model

Use Model for Prediction

Figure 2.5: Supervised learning work-flow

2. Unsupervised Learning

Unsupervised learning is a technique for finding hidden patterns or intrinsic structures in the data. It is used to draw inferences from datasets, which consists of input data without labelled responses. Clustering is a common technique for unsupervised learning, which aims at finding patterns or similarities using the Euclidean or the probabilistic distance. This technique is widely used in bioinformatics, computer vision, and medical imaging. Some representative algorithms for clustering are listed in Figure 2.4. A workflow model of unsupervised learning is also explained in Figure 2.6. A process of model generation starts with feature extraction from unlabelled data.

2.4.1. Machine Learning Classification 35 Extracted features are then used to group subsets of data into clusters and later on used to predict labels of new data.

Raw Data

(Unlabelled)

Feature

Extraction

Group into

Classes/

Clusters

Predict Label

Figure 2.6: Unsupervised learning workflow

3. Semi-Supervised Learning

Semi-supervised learning falls between the supervised and the unsupervised learning, because it considers labelled as well as unlabelled data for training. Some of the semi-supervised learning models are self-training models, mixture models, co-training and multi-view learn- ing, graph-based methods, and semi-supervised support vector machines. Furthermore, since semi-supervised learning makes assumptions about the distribution of unlabelled data, for ex- ample smoothness assumption, cluster assumption, and manifold assumption, so it is a form of transductive and inductive learning.

A workflow model of semi-supervised learning is explained in Figure 2.7. A process of model generation starts with feature extraction from few labelled and a large amount of unlabelled data. Extracted features are then used to fit a model based on initial assumptions about the underlying relationship of features with the response variable. Learned model is then used to infer correct labels/values for the unlabeled data and tested on test data set. If the model is performing well on test data then the model is further used to predict future values. Most of the learning algorithms defined for supervised and unsupervised learning can also be used under semi-supervised learning approach.

4. Reinforcement Learning (RL)

RL aims for mapping situations to the actions so as to maximize the reward. RL is similar to behavioural psychology. Specifically, RL is a trial and error learning technique, which aims to strike a reasonable trade-off between exploration and exploitation- exploitation is when an agent

Raw Data (Few labelled,

mostly unlabelled)

Feature

Extraction Fit a Model

Train to infer correct label Test a Model Use Model for Prediction

Figure 2.7: Semi-supervised learning workflow

uses its past experience, while exploration is to make a better action selection in future. On the success of an action, the agent receives a numerical reward. Hence, the agent tries to learn and select those actions, which would maximize its accumulated reward over time. A workflow model of RL is explained in Figure 2.8. Agent learns to react to some action by trial and error process of applying actions on the environment and analysing the result of action by reviewing next state as well as reward value. This type of learning is mostly applied in the robotics field. Some commonly known algorithms under this category are listed in Figure 2.4

Agent

Environment

Action

Reward

State

Figure 2.8: Reinforcement learning work-flow