Additional insight into this database is provided in this note from NASA
Ames' John Stutz to UCI's John Gennari:

======================================================================
Date: Wed 25 Apr 90 18:27:22-PST
From: JOHN STUTZ <STUTZ@PLUTO.ARC.NASA.GOV>
To: gennari@ICS.UCI.EDU
Subject: RE: [Resent-MM Mail]   Astral data.

John

    There was an error in the IRAS-LRS data documentation.  We ignored
the presence of 5 fields of scaling information used to decompress the
original data.  This probably occurred because, once we had done the
decompression, we never again considered these fields.  The corrected
documentation has been sent to Dave Aha and placed in the database.  I
include the changes below.

    With respect to the choice of attributes for clustering, the
information is all in the location and spectral intensity values.  The
'standard IRAS-LRC classification was based on spectral features.  These
include the peaks, plateaus and valleys, and their locations and
relative magnitudes.  They also considered the trend in intensity at
each end of each spectra.  If you could extract these features, you
might expect to find classes fairly close to the official ones.  But
that could be a fairly complex modeling process that would be very
specific to the field of spectral analysis.  The plain truth is that
this is not a good database for making predictions from a small set of
features.  AutoClass did well, despite it's vastly oversimplified
probability model, because it used all 93 spectral attributes and 10
times as many instances.

    I doubt you can get anything interesting from the sky locations
alone.  However, the greatest variations in our classes were in the
region of 8 to 11 microns.  So this is a good region to concentrate on.
You might also look for variations of the ratio of the higher and lower
wavelength intensities.  Given that the high magnitude declination
sources mostly represent those which are relatively close, you might
find declination dependencies.  AutoClass found several pairs in which
the two classes had very similar class mean spectra and quite different
declination distributions.  Since we did not use the location attributes
there must be a relation between spectrum and position.

   Wishing you the best of success with your work.

   John Stutz

- -------------------------------------------------------------------
4. Relevant Information:

    The Infra-Red Astronomy Satellite (IRAS) was the first attempt to
map the full sky at infra-red wavelengths.  This could not be done from
ground observatories because large portions of the infra-red spectrum is
absorbed by the atmosphere.  The primary observing program was the full
high resolution sky mapping performed by scanning at 4 frequencies. The
Low Resolution Observation (IRAS-LRS) program observed high intensity
sources over two continuous spectral bands.  This database derives from
a subset of the higher quality LRS observations taken between 12h and
24h right ascension. 

    This database contains 531 high quality spectra derived from the
IRAS-LRS database.  The original data contained 100 spectral
measurements in each of two overlapping bands.  Of these, 44 blue band
and 49 red band channels contain usable flux measurements.  Only these
are included here.  The original spectral intensities values are
compressed to 4-digits, and each spectrum includes 5 rescaling
parameters.  We have used the LRS specified algorithm to rescale these
to units of spectral intensity (Janskys).  Total intensity differences
have been eliminated by normalizing each spectrum to a mean value of
5000.
	
    This database was originally obtained for use in development and
testing of our AutoClass system for Bayesian classification.  We have
not retained any results from this development, having concentrated our
efforts of a 5425 element version of the same data.  Our classifications
were based upon simultaneous modeling of all 93 spectral intensities.
With the larger database we were able to find classes that correspond
well with known spectral types associated with particular stellar types.
We also found classes that match with the spectra expected of certain
stellar processes under investigation by Ames astronomers.  These
classes have considerably enlarged the set of stars being investigated by
those researchers.  

5. Number of Instances: 531

6. Number of Attributes: 103 (including the 10-attribute "header")

7. Attribute Information: 
    1. LRS-name: (Suspected format: 5 digits, "+" or "-", 4 digits)
    2. LRS-class: integer - The LRS-class values range from 0 - 99 with
	the 10's digit giving the basic class and the 1's digit giving
	the subclass. These classes are based on features (peaks,
	valleys, and trends) of the spectral curves.  
    3. ID-type: integer
    4. Right-Ascension: float - Astronomical longitude. 1h = 15deg
    5. Declination: float - Astronomical lattitude. -90 <= Dec <= 90
    6. Scale Factor: float - Proportional to source strength
    7. Blue base 1: integer - linear rescaling coefficient
    8. Blue base 2: integer - linear rescaling coefficient
    9. Red base 1: integer - linear rescaling coefficient
   10. Red base 2: integer - linear rescaling coefficient
   11-54: fluxes from the following 44 blue-band channel wavelengths: 
   55-103: fluxes from the following 49 red-band channel wavelengths: 

