---
type: learning
area: learning
status: learning
date: 2026-05-03
created: 2026-05-03
updated: 1980-01-01
tags:
  - learning
---
### Plan
![[img-20240927-120555-7b979306.png]]
### How SOMs work?
![[img-20240927-120706-add833b1.png]]

It's an unsupervised algorithm
Used to reduce dimensionality (reduce amount of columns)
For example where we can use this technique:

![[img-20240927-121904-787ef945.png]]

### How do SOMs learn?
All data passed to self organizing maps should be normalised or standardised
No activation function is needed and weights are coordinate of nodes in our dimensions
### Example
![[img-20240927-200901-e15c5c59.png]]
Calculating between all nodes for all rows of data
and after all calculations we will update the weights of most close node (omega)
![[img-20240927-201232-95e71692.png]]
- After we will update weights of all nodes that are close to this node (omega) and this operation of update is based on while the node is most close to (omega) will be updated more.
- Radius is taken on the start as large value
- Radius became much smaller with time

Notes:
![[img-20240927-202929-afb77d3b.png]]

Random Initialization Trap
we fix the issue of every time selection centro-ids randomly with k-means ++

### How to input number of clusters to k-means:
Based on WCSS value:
- ![[img-20240927-212038-dbf8ba77.png]]
- When increasing number of clusters -> WCSS value decrease

The Optimal value of WCSS can b identified from chart like that (elbow method)
![[img-20240927-212334-6a5a5e11.png]]

### Summary
![[img-20240927-222132-ff340aa6.png]]