Top Page | Upper Page | Contents | About This Site | JAPANESE

Independent Component Analysis

Independent Component Analysis (ICA) is younger than Principal Component Analysis (PCA) and Factor Analysis (FA).

Cocktail Party Problem

If there are some speakers, each voice data in the room is mixed data of each speaker.

"Cocktail Party Problem" is the problem to get the data of each speaker from the mixed data.

ICA is developed as the solution of this problem


"In1" and "In2" are the mixed data of "Fact1" and "Fact2". I tried ICA with this set of data by RapidMiner.

Outputs are "Out1" and "Out2". Outputs are similar to Fact1 and Fact2.

As Method to Outliers

Some peple introduce ICA as the filter of outlier or abnormal data.

Histgram of the data of In2 is below. It is not so easy to filter this case.

Difference from Other Methods

Principal Component Analysis (PCA)

Example by PCA is below. Outputs are not similar to Facts.

PCA is the method to analyze main movement of the data set.

Factor Analysis (FA)

The purpse of ICA is "To finr factors". This is same to FA.

To Treat Time

The data of sound is time depended. But many mehtods in Multi-Variable Analysis and Data Mining do not take care about the infomation of time in the data.

The logic of ICA also does not care about time. But many people studies to use the infomation of time by ICA. This studies are useful for other methods.

Example of this page does not use the infomation of time. But ICA works well.

NEXT Item Response Theory