Classifications of ML Methods

Here are a few ways commonly used to classify ML systems:

Supervision

In this classification, there are four main classes based on the amount or type of supervision

Data
- Labelled examples \large{({\bold x_i},y_i)_{i=1}^N}
- \large x_i is a feature vector(\large n-dimensional vector of numerical features \large x^d)
  - represent objects numerically e.g. for an image, \large x^{(1)} could be the hue, and \large x^{(2)} could be the intensity
  - useful for comparing objects e.g. euclidean distance
  - granularity depends on the purpose
- \large y_i can take the form of a member of a set, real number, matrix, vector etc.
Tasks
- Classification e.g. spam filter
- Predict numeric values based on predictors (features) e.g. house price given room numbers and areas
Goal
- Train a model on a dataset to predict labels based on input feature vectors
Methods
- Linear regression
- Logistic regression
- kNN
- SVM
- DT (& random forests)
- Neural networks

Data
- Unlabelled examples \large{({\bold x_i})_{i=1}^N}
- \large x_i is a feature vector
Tasks
- Clustering - group data points with shared attributes to extrapolate a relationship e.g. molecular structure similarity
- Anomaly/Outlier detection - find outliers e.g. fraud-detection
- Rule learning
Goal
- Clustering - transform feature vector \large \bold x into a useful value e.g. an id
- Dimensionality reduction - output feature vector should have fewer features than \large \bold x
- Anomaly/Outlier detection - output value quantifies the difference of \large \bold x from the data
Methods
- Clustering
  - Exclusive
    - K-means
    - Hierarchical
  - Soft - more probabilistic
    - GMM
    - Expectation Maximisation
- Association
  - Apriori
  - Eclat
- Dimensionality reduction
  - PCA
  - SVD
  - LLE
  - t-SNE
Challenges
- Computation and time complexity of training
- Can be unclear as to basis for clustering
- Accuracy

Data usually partially labelled
Combination of supervised and unsupervised
Methods
- Deep Belief Networks (DBN) - based on stacked Restricted Boltzmann machines

Agent - learning system
- observes the environment - state
- makes decisions
- performs actions
- feedback loop - penalties or rewards - aims to maximise rewards
Learns a policy - a function that takes the state as an input feature vector and outputs an action that leads to the highest expected average reward

Split into batch (non-incremental) and online

Offline - unable to learn incrementally from a data stream (usually due to complexity)
System trained first then applied, without learning further unless it is taken offline and retrained
Can be automated e.g. weekly
High computational requirements

Training using small sequential chunks of data - streamed
Does not require storage of previous data
Learning steps are of low complexity therefore can be performed on mobile systems
Can also be used to process extremely large datasets as a stream
Learning rate
- if high, will forget old data faster, sensitive to noise
- if low, inertia will be high, slower to learn, less sensitive to noise or outliers
Corrupted or errors in the stream can affect the performance in real-time
- pair with anomaly detection

Can also categorise based on models of generalisation

Learn from prior examples, then generalise to new data using a measure of similarity

Build or select a model from prior examples
Define a fitness function or cost function
Minimise cost or maximise fitness, depending on the chosen model
Train the model on the training data to optimise parameters for a reasonable fit
Make predictions

Burkov, A. (2019) The Hundred-Page Machine Learning Book.
Geron, A. (2017) Hands-On Machine Learning with Scikit-Learn & TensorFlow : concepts, tools, and techniques to build intelligent systems.