Components

  • Unsupervised Learning: SOMs don’t require labeled data → learns patterns and structure from raw data
  • Competitive Learning: Neurons compete to represent input data
    • winning neuron and its neighbors update their weights to better match the input
  • Topology Preservation: Similar data points are mapped to nearby neurons, making SOMs excellent for visualizing clusters and relationships

How it Works

  1. Initialization: Each neuron in the grid starts with random weights
  2. Best Matching Unit (BMU): For each input, the neuron whose weights are closest to the input is selected
  3. Weight Update: The BMU and its neighbors adjust their weights to better match the input
  4. Decay: Learning rate and neighborhood size shrink over time for convergence

Goal

  • Clustering - Grouping similar data points
  • Dimensionality Reduction - Visualizing complex datasets in 2D
  • Anomaly Detection
  • Utilize competitive learning to map high-dimensional input data onto a typically 2D grid
    • Preserves topological relationships in the process
      • Similar inputs are grouped closer together in the map
      • Dissimilar ones are placed farther apart
    • Crucial in exploratory data analysis

Examples

  • Businesses
    • Segment customers based on
      • Purchasing patterns
      • Demographic attributes
      • Preferences
  • Retail analytics (customer segmentation and personalization)
    • Often collect extensive, unlabeled transactional and behavioral data from customers
      • Require advanced analytical methods for actionable insights
    • Use autoencoders
      • effectively compress transaction histories and browsing data into concise representations
  • Following dimensionality reduction
    • Further categorize customers into distinct segments
    • Facilitates the
      • Precise targeting of marketing campaigns
      • Personalized product recommendations
      • Efficient inventory management