Unsupervised Learning

A centroid-based algorithm that partitions data into k distinct, non-overlapping clusters by minimizing the variance within each group.

• Mechanism: Iteratively assigns data points to the nearest centroid, then recalculates centroids as the mean of assigned points.
• Network Use Case: Grouping cells with similar traffic load patterns to define baseline performance profiles.
• Key Limitation: Requires pre-specifying the number of clusters (k) and assumes spherical cluster shapes.
• Anomaly Connection: Data points far from any centroid are flagged as potential anomalies.

A density-based algorithm that groups closely packed data points and identifies points in low-density regions as outliers or anomalies.

• Core Parameters: Epsilon (ε) defines the neighborhood radius; MinPts defines the minimum points to form a dense region.
• Arbitrary Shapes: Unlike K-Means, DBSCAN can discover clusters of arbitrary shape and does not require specifying the number of clusters upfront.
• Network Use Case: Identifying micro-regions of high interference or unusual handover patterns in cellular topologies.
• Key Advantage: Inherently classifies noise points, making it a dual-purpose clustering and anomaly detection algorithm.

A probabilistic model that assumes all data points are generated from a mixture of a finite number of Gaussian distributions with unknown parameters.

• Soft Clustering: Unlike K-Means, GMM provides a probability score for each point belonging to each cluster, quantifying membership uncertainty.
• Expectation-Maximization (EM): The algorithm iteratively estimates the mean, covariance, and mixing coefficients of each Gaussian component.
• Network Use Case: Modeling the distribution of latency measurements where overlapping performance states exist, such as 'normal', 'degraded', and 'congested'.
• Anomaly Connection: Points with very low probability density under the learned model are flagged as anomalies.

A method that builds a tree-like structure (dendrogram) of clusters by either successively merging smaller clusters (agglomerative) or splitting larger ones (divisive).

• No Predefined K: Does not require specifying the number of clusters in advance; the dendrogram can be cut at any level.
• Linkage Criteria: Defines how distance between clusters is calculated—single, complete, average, or Ward's method.
• Network Use Case: Creating a taxonomy of network fault signatures by grouping similar alarm sequences.
• Visual Insight: The dendrogram provides an intuitive visualization of data structure at multiple granularities.

A dimensionality reduction technique that projects high-dimensional telemetry data onto a lower-dimensional subspace while preserving maximum variance.

• Mechanism: Computes eigenvectors (principal components) of the data's covariance matrix, ordered by explained variance.
• Reconstruction Error: Anomalies are detected by projecting data into the reduced space and back; points with high reconstruction error deviate from the normal subspace.
• Network Use Case: Reducing hundreds of correlated PM counters to a handful of uncorrelated features that capture the essential system state.
• Key Insight: The first few principal components often represent the dominant operational modes of the network.

A neural network trained to copy its input to its output through a bottleneck layer, learning a compressed latent representation of the data.

• Architecture: Consists of an encoder that compresses input and a decoder that reconstructs it. The bottleneck forces the network to learn only the most salient features.
• Anomaly Detection via Reconstruction Error: Trained exclusively on normal data, the autoencoder will produce high reconstruction error for anomalous inputs it cannot faithfully reproduce.
• Network Use Case: Learning the complex, non-linear normal behavior of a cell site's full KPIs to detect subtle degradation preceding a fault.
• Key Advantage: Captures non-linear relationships that linear methods like PCA miss.

The most common unsupervised learning task, where algorithms partition data points into groups (clusters) based on inherent similarity. Unlike classification, no predefined labels exist.

• K-Means: Partitions data into k distinct clusters by minimizing intra-cluster variance.
• Hierarchical Clustering: Builds a tree of clusters (dendrogram) without requiring a pre-specified number of groups.
• Density-Based Clustering (DBSCAN): Forms clusters based on dense regions of data points, effectively identifying noise and arbitrary-shaped clusters.

Techniques that reduce the number of random variables under consideration by transforming high-dimensional data into a lower-dimensional manifold, preserving its essential structure.

• Principal Component Analysis (PCA): A linear method that finds orthogonal axes (principal components) maximizing variance.
• t-SNE: A non-linear technique optimized for visualizing high-dimensional data in 2D or 3D space by preserving local similarities.
• Autoencoders: Neural networks trained to compress data into a latent-space representation and reconstruct it, learning a non-linear generalization of PCA.

A rule-based method for discovering interesting relations between variables in large databases. It identifies frequent if-then patterns without a specific target variable.

• Market Basket Analysis: The classic application, finding products frequently bought together (e.g., bread and butter).
• Apriori Algorithm: Identifies frequent individual items and extends them to larger item sets, pruning infrequent combinations.
• Support & Confidence: Key metrics measuring how often a rule appears in the dataset and how reliable the implication is.

A branch of unsupervised learning focused on learning the true probability distribution of the training data to generate new, synthetic data points with similar characteristics.

• Generative Adversarial Networks (GANs): Pit a generator against a discriminator in a zero-sum game to produce highly realistic synthetic data.
• Variational Autoencoders (VAEs): Learn a probabilistic latent space, enabling smooth interpolation and controlled generation of new samples.
• Diffusion Models: Learn to reverse a gradual noising process, generating high-fidelity data by denoising random noise step-by-step.

A paradigm where a supervised learning signal is created automatically from the unlabeled data itself, often by masking parts of the input and training the model to predict the missing content.

• Masked Language Modeling (MLM): Used in BERT, where random words in a sentence are masked and the model learns to predict them from context.
• Contrastive Learning: Learns representations by pulling similar (positive) pairs closer and pushing dissimilar (negative) pairs apart in the embedding space.
• Pretext Tasks: Surrogate tasks like rotation prediction or jigsaw puzzle solving that force the model to learn semantic features.

The identification of rare items, events, or observations that raise suspicions by differing significantly from the majority of the data. Often framed as an unsupervised problem when labeled anomalies are scarce.

• Isolation Forest: Explicitly isolates anomalies by randomly partitioning data, exploiting the fact that anomalies are 'few and different' and require fewer splits.
• One-Class SVM: Learns a tight boundary around the 'normal' data distribution, classifying any point outside this boundary as an anomaly.
• Reconstruction Error: Autoencoders trained only on normal data will have a high reconstruction error when attempting to reconstruct an anomalous sample.