Machine Learning Week5 Unsupervised Learning.pdf

Unsupervised Learning in Machine Learning Unsupervised learning is a powerful branch of machine learning. It finds hidden patterns in data without labeled responses. This approach is crucial for extracting insights from complex, unlabeled datasets. Key areas include clustering, dimensionality redu...

Unsupervised Learning in Machine Learning Unsupervised learning is a powerful branch of machine learning. It finds hidden patterns in data without labeled responses. This approach is crucial for extracting insights from complex, unlabeled datasets. Key areas include clustering, dimensionality reduction, and association rules. These techniques reveal underlying structures and relationships in data. by bien medina Introduction to Unsupervised Learning 1 Data-Driven Discovery 2 Versatile Applications Unsupervised learning Used in various fields like algorithms explore data customer segmentation, without predefined labels. anomaly detection, and They identify inherent feature learning. Adapts to structures and patterns diverse datasets and problem autonomously. domains. 3 Complementary to Supervised Learning Often used as a preprocessing step. Can improve supervised learning models by revealing hidden data characteristics. Clustering Algorithms 1 Partitioning Methods Divide data into non-overlapping subsets. K-means is a popular example in this category. 2 Hierarchical Methods Create a tree-like structure of clusters. Can be agglomerative (bottom-up) or divisive (top-down). 3 Density-Based Methods Form clusters based on areas of high data point density. DBSCAN is a well-known algorithm. K-Means Clustering Initialize Centroids Randomly select K points as initial cluster centers. K is predetermined by the user. Assign Points Assign each data point to the nearest centroid based on Euclidean distance. Update Centroids Recalculate centroids as the mean of all points in each cluster. Iterate Repeat steps 2-3 until convergence or maximum iterations reached. Hierarchical Clustering Agglomerative Clustering Divisive Clustering Linkage Criteria Bottom-up approach. Starts with Top-down approach. Begins with all data Methods to determine cluster similarity: individual data points as clusters. in one cluster. Recursively splits clusters single-linkage, complete-linkage, Progressively merges closest clusters until until each point is its own cluster. average-linkage, and Ward's method. one cluster remains. Density-Based Clustering DBSCAN Algorithm Flexible Cluster Shapes Density-Based Spatial Clustering Can discover clusters of arbitrary of Applications with Noise. Forms shapes. Not limited to convex clusters in high-density regions, clusters like K-means. identifies outliers in low-density areas. N oise Handling Parameter Sensitivity Explicitly labels outliers as noise. Requires careful tuning of Robust to outliers and can epsilon (neighborhood radius) handle datasets with varying and minPoints parameters for densities. optimal results. Dimensionality Reduction Techniques Linear Methods Feature Selection PCA, LDA, and Factor Analysis. Preserve global structure through linear Identify most relevant features. Techniques include correlation-based and transformations. mutual information approaches. 1 2 3 Non-linear Methods t-SNE, UMAP, and Isomap. Capture complex, non-linear relationships in high- dimensional data. Principal Component Analysis (PCA) Dimensionality Reduction Variance Explained Projects high-dimensional data onto Quantifies information retained by each lower-dimensional subspace. Preserves principal component. Aids in choosing maximum variance. number of components. Feature Extraction N oise Reduction Creates new features as linear Lower-order components often capture combinations of original features. Useful noise. Discarding them can denoise for data compression. data. t-SNE Full Name t-Distributed Stochastic Neighbor Embedding Approach Non-linear dimensionality reduction Strength Preserves local structure and reveals clusters Weakness Computationally expensive, non- deterministic Best For Visualizing high-dimensional data in 2D or 3D Association Rule Mining 1 Market Basket Analysis 2 Apriori Algorithm Discovers relationships Popular method for finding between items in transaction frequent itemsets. Uses a data. Identifies frequently co- bottom-up approach with occurring items. candidate generation. 3 FP-Growth Algorithm 4 Evaluation Metrics Efficient alternative to Apriori. Support, confidence, and lift Uses a compact data measure rule strength and structure called FP-tree. significance. Guide rule selection. Applications of Unsupervised Learning Unsupervised learning, a powerful branch of machine learning, has revolutionized the way we extract insights from data without explicit guidance. This fascinating field allows algorithms to discover hidden patterns and structures within vast datasets, opening up a world of possibilities across various industries and applications. In this presentation, we'll explore the diverse applications of unsupervised learning, from clustering algorithms to generative models, and examine how these techniques are transforming business, science, and technology. Join us on a journey through the exciting landscape of unsupervised learning and discover its potential to unlock valuable insights from complex, unlabeled data. by bien medina Clustering Algorithms 1 K-Means Clustering 2 Hierarchical Clustering K-means is a popular algorithm that This method creates a tree-like partitions data into K distinct, non- structure of clusters, allowing for multi- overlapping clusters. It's widely used in level grouping. It's valuable in customer segmentation, image bioinformatics for gene expression compression, and anomaly detection. analysis and in marketing for customer hierarchy understanding. 3 DBSCAN 4 Gaussian Mixture Models Density-Based Spatial Clustering of GMMs model data as a mixture of Applications with Noise is effective for Gaussian distributions, making them discovering clusters of arbitrary shape. suitable for complex clustering tasks It's particularly useful in spatial data such as speaker identification and analysis and outlier detection in image segmentation in computer financial transactions. vision. Dimensionality Reduction 1 Principal Component Analysis (PCA) PCA is a cornerstone technique that reduces data dimensionality while preserving maximum variance. It's extensively used in facial recognition, finance for portfolio optimization, and in bioinformatics for gene expression analysis. 2 t-SNE (t-Distributed Stochastic Neighbor Embedding) t-SNE is particularly effective for visualizing high-dimensional data in 2D or 3D space. It's widely applied in single-cell RNA sequencing data analysis and for visualizing complex datasets in machine learning research. 3 Autoencoders These neural network-based models learn compact representations of data. They're used in image and speech denoising, anomaly detection in manufacturing, and generating realistic synthetic data for training other models. 4 UMAP (Uniform Manifold Approximation and Projection) UMAP is a more recent technique that often outperforms t-SNE in preserving both local and global structure. It's gaining popularity in bioinformatics, particularly in single-cell genomics and proteomics data analysis. Anomaly Detection Financial Fraud Detection Network Intrusion Detection ManufacturingQuality Control Unsupervised learning algorithms like In cybersecurity, anomaly detection plays Unsupervised learning methods are Isolation Forests and One-Class SVMs are a vital role in identifying potential threats. employed in industrial settings to detect crucial in identifying fraudulent Techniques such as autoencoders and defects in products. By analyzing sensor transactions. These methods can detect clustering algorithms can spot unusual data and images, algorithms can identify unusual patterns in credit card usage, network traffic patterns, helping to anomalies in production lines, ensuring insurance claims, and stock market prevent data breaches and cyberattacks high-quality output and reducing waste in trading, helping financial institutions before they cause significant damage. manufacturing processes. safeguard their customers and assets. Recommendation Systems Content-Based Filtering Collaborative Filtering This approach recommends items similar to By identifying patterns in user behavior, this those a user has liked in the past. It's widely method recommends items based on similar used in movie and book recommendation users' preferences. It's the backbone of many systems, analyzing features like genre, actors, e-commerce and social media or authors to suggest new content. recommendation engines. Hybrid Systems Graph-Based Methods Combining content-based and collaborative These techniques use network structures to filtering, hybrid systems offer more accurate model relationships between users and items. and diverse recommendations. They're used They're particularly effective in social network- by streaming services like Netflix to provide based recommendations and knowledge personalized viewing suggestions. graph applications. Image Segmentation Data Preparation Feature Extraction Clustering Post-processing Raw images are preprocessed, Unsupervised techniques like Algorithms such as K-means or The segmented image is refined using including resizing, normalization, and autoencoders or principal component DBSCAN group similar pixels or techniques like morphological augmentation to enhance the analysis extract meaningful features regions together based on color, operations or conditional random dataset's diversity and quality. from the images, reducing texture, or other extracted features. fields to improve boundary accuracy dimensionality while preserving and overall segmentation quality. important information. Topic Modeling Algorithm Description Applications Latent Dirichlet Probabilistic model that Content Allocation (LDA) discovers topics in recommendation, documents document clustering Non-Negative Matrix Factorizes document- Text summarization, Factorization (NMF) term matrix into topic- feature extraction term and document- topic matrices Pachinko Allocation Extends LDA to model Hierarchical topic Model (PAM) topic correlations discovery, improved coherence Dynamic Topic Models Captures topic evolution Trend analysis, historical over time document studies Generative Models Variational Autoencoders (VAEs) Generative Adversarial Networks (GANs) VAEs learn to encode data into a latent space and then reconstruct it, allowing for GANs consist of a generator and generation of new samples. They're used in discriminator network competing against image generation, drug discovery, and each other. They excel in creating highly creating synthetic data for privacy- realistic images, videos, and even music. preserving machine learning. Applications include deepfakes, art creation, and data augmentation for training other models. Autoregressive Models Flow-based Models These models, like GPT for text or PixelCNN Normalizing flows learn invertible for images, generate data sequentially. transformations between simple They're powerful in language modeling, text distributions and complex data. They're completion, and creating coherent long- used in density estimation, anomaly form content. detection, and generating molecular structures in drug discovery. Unsupervised Representation Learning 1 Self-Supervised Learning This approach creates supervised tasks from unlabeled data. For example, predicting the next word in a sentence or rotating an image. It's revolutionizing NLP and computer vision by learning rich representations from vast amounts of unlabeled data. 2 Contrastive Learning Techniques like SimCLR learn representations by contrasting similar and dissimilar samples. This has led to state-of-the-art results in image classification and transfer learning, especially when labeled data is scarce. 3 Deep Clustering Methods like DeepCluster combine representation learning with clustering. They iteratively cluster the data and use the cluster assignments as pseudo-labels for representation learning, improving both clustering and feature extraction. 4 Energy-Based Models These models learn to assign low energy to observed data and high energy to other configurations. They're particularly useful in learning complex data distributions and have applications in anomaly detection and generative modeling. Applications in Business and Industry Retail and E- commerce Manufacturing and Industry 4.0 Healthcare and Life Sciences Unsupervised learning powers personalized In smart factories, anomaly detection Unsupervised learning aids in drug discovery by product recommendations, dynamic pricing algorithms monitor equipment health for identifying potential compounds. It's used in strategies, and inventory optimization. Retailers predictive maintenance. Unsupervised learning medical imaging for anomaly detection in X-rays use clustering for customer segmentation and also optimizes supply chains, improves quality or MRIs. In genomics, it helps in understanding market basket analysis to understand control through image segmentation, and gene expression patterns and discovering new purchasing patterns and improve store layouts. enhances process efficiency through pattern subtypes of diseases. discovery in sensor data.