Comparison of Supervised, Unsupervised, Semi-Supervised and Reinforcement
Introduction
Supervised, unsupervised learning, semi-supervised and reinforced learning are 4 fundamental approaches of machine learning:
- Supervised Learning Builds a model based labelled data.
- Unsupervised Learning Builds a model based on a unlabelled data.
- Semi-Supervised Learning Builds a model based on a mix of labelled and unlabelled data. This sits between supervised and unsupervised learning approaches.
- Reinforcement Learning This is a feedback-based learning method, based on a system of rewards and punishments for correct and incorrect actions respectively. The aim is for the “learning agent” to receive maximum reward and hence improve its performance.
Overview comparison between these methods
| Category | Supervised | Unsupervised | Semi-supervised | Reinforcement |
|---|---|---|---|---|
| Input data | All data is labelled | All data is unlabelled | Partially labelled | No predefined data |
| Training? | External supervision | No supervision | (External supervision) | No supervision |
| Use | Calculate outcomes | Discover underlying patterns | Improve learning performance | Learn a series of outcomes |
| Computational complexity | Simple | Complex | Depends | Complex |
| Accuracy | Higher | Lesser | Lesser | Good for trial/error situations |
Example algorithms under each approach
Below is a basic comparison table of the different approaches with a few example algorithms:
| Supervised | Unsupervised | Semi-supervised | Reinforcement |
|---|---|---|---|
| Decision trees | K-means | Generative adversial networks | Q-learning |
| Support Vector Machine | A-priori | Self-trained Naïve Bayes classifier | Model-based value estimation |
| Linear regression | Hierarchical clustering | Low-density separation | Policy optimization |
| Logistic regression | K Nearest Neighbours | Laplacian regulation | State-Action-Reward-State-Action |
| Naive Bayes | Principal Component Analysis | Heuristic approaches | Deep Q Network |
Example uses of each approach
Below is a basic comparison table of the different approaches with a few example uses:
| Supervised | Unsupervised | Semi-supervised | Reinforcement |
|---|---|---|---|
| Image recognition | Customer segmentation | Text document classifier | Playing games e.g. chess game |
| Market prediction e.g. house prices | Anomaly e.g. fraud detection | Speech analysis | Self-driving cars |
Challenges using each approach
| Supervised | Unsupervised | Semi-supervised | Reinforcement |
|---|---|---|---|
| Pre-processing of data may be time consuming | More time required by user e.g. for interpretation | Complex iterative process | Choosing reward structures wisely |
| Cannot give “unkown” information as per unsupervised learning | May result in less accurate predictions compared to supervised learning | Not as accurate as supervised learning | Fast learing given small samples |
| Cannot handle “complex tasks” | Computationally more complex that supervised learning | Cannot handle more “complex tasks” | Not preferable for learning “simple problems” |