Unsupervised learning

Unsupervised learning - overview

Please do not copy without permission. © ExploreAI 2024.
Unsupervised learning

Historical context of unsupervised learning

Unsupervised learning has evolved through decades of The advent of computers and algorithmic
research and innovation. Initially, the focus in data development marked a pivotal shift, enabling the
analysis was predominantly on manual classification exploration of automated data grouping without
and data interpretation. predefined labels.

Stuart Lloyd's implementation of _DBSCAN Density-based spatial _Internet and computing advances _
The _k-means clustering algorithm_ clustering of applications with noise) have greatly improved, enhancing
for _pulse-code modulation_, introduced to manage data in algorithms' ability to _handle_
published in 1982. _densely populated regions._ _complex datasets._

1957 1963 1973 1980 1990s 2000s

Non-linear
The Hierarchical Density-based dimensionality … Deep learning
beginnings clustering clustering reduction to present influence

Hierarchical clustering introduced, Introduction of methods such as Introduction of deep learning,

enabling the visualisation of_ _t-SNE_ (t-Distributed Stochastic including autoencoders , transformed
_tdata relationships_ through Neighbour Embedding) for unsupervised learning by uncovering_
_tree-like models._ _visualisation high-dimensional_ data. _ hidden patterns in unlabelled data._

2
Unsupervised learning

Learning paradigms in machine learning

|
Machine learning can be broadly categorised into three main paradigms, each with unique
methods and applications. Understanding these differences is crucial for choosing the right
approach for specific data science tasks.

Learning paradigm Definition Data requirement Use cases

Learn a model from labelled training Requires a dataset with Image recognition, spam detection,
Supervised learning
data to make predictions or decisions. input-output pairs. regression tasks.

Unsupervised Explore the underlying structure or No labels are needed, only the Customer segmentation, anomaly

learning distribution in data without labels. input data. detection, association mining.

Learn to make decisions by

Reinforcement No predefined labels, but Game AI, robotics, real-time
performing actions and receiving
learning requires a reward system. decisions.
rewards or penalties.

3
Unsupervised learning

Challenges in unsupervised learning

| Understanding the unique challenges that unsupervised learning presents is crucial for applying
itʼs techniques successfully.

Model ● Unsupervised learning does not use labelled data, making it difficult to confirm whether the identified
patterns are meaningful or just noise.
validation
● Interpreting cluster meanings demands domain knowledge, often complicating direct interpretations.
without ● Use measures like the Silhouette Score to evaluate cluster validity, emphasising their importance in
labels confirming the practical significance of the findings.

Determining ● One of the fundamental decisions in clustering involves determining the number of clusters (k).
● Choosing too few clusters might oversimplify the model, while too many can overfit the data.
the number ● Introduce metrics like the Elbow Method, Silhouette Score, and Gap Statistic that help infer the optimal
of clusters number of clusters.

● Higher dimensions can make clustering exponentially harder, a phenomenon known as the curse of
Handling high dimensionality.
dimensionality ● Dimensionality reduction techniques like PCA Principal Component Analysis) simplify the data without
losing critical information.

4
Unsupervised learning

Applications and industry relevance

● Algorithms like k-means clustering are used to segment patients based on various characteristics (e.g.,
Healthcare symptoms, demographics).
● Impact: This segmentation helps in personalised treatment plans and resource allocation.

● Techniques such as anomaly detection are employed to identify unusual patterns in financial
Finance transactions which may indicate fraudulent activities.
● Impact: Increases security by early detection of fraud, saving millions in potential losses.

● Retail companies use clustering to group customers based on purchasing behaviour and preferences to
Marketing target marketing efforts more effectively.
● Impact: Optimises marketing strategies, enhancing customer engagement and boosting sales.

● Clustering algorithms analyse sensor data from manufacturing equipment to identify patterns and
Manufacturing optimise processes without predefined labels.
● Impact: Improves operational efficiency and product quality, reducing costs and waste.

● Unsupervised learning is used to monitor network traffic and spot unusual patterns that could indicate a
Cybersecurity security breach.
● Impact: Enhances network security by proactively identifying and mitigating risks.
5
Unsupervised learning

Future directions in unsupervised learning

Deep learning Autonomous decision-making
Deep unsupervised learning: Combining deep learning with
Self-learning systems: Future unsupervised learning models are
unsupervised methods to create powerful models capable of
being developed to not only analyse data but also make
learning complex, high-dimensional data without labelled
decisions based on identified patterns without human
examples.
intervention.
Example: Autoencoders and generative models are learning
Example: Autonomous vehicles use unsupervised learning to
architectures used to understand and generate new data based
interpret real-time data and make immediate driving decisions.
purely on its inherent structure and distribution.

Complex data patterns Other AI domains

Complex data patterns: With advancements in algorithmic AI Integration: Unsupervised learning is increasingly being used
techniques, unsupervised learning is now better at identifying alongside supervised learning, reinforcement learning, and other
subtler and more complex patterns in data. AI techniques to create more robust and adaptable AI systems.

Applications: From detecting anomalies in financial transactions Potential impact: This convergence is expected to unlock
to understanding genetic sequences in bioinformatics, the unprecedented capabilities in AI, from improving learning
capabilities of unsupervised learning in pattern recognition are efficiency to enabling machines to understand and interact with
vast and growing. the world in fundamentally new ways.
6