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
- Initialization: Each neuron in the grid starts with random weights
- Best Matching Unit (BMU): For each input, the neuron whose weights are closest to the input is selected
- Weight Update: The BMU and its neighbors adjust their weights to better match the input
- 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