Johns Hopkins University Ionosphere database

1. Title: Johns Hopkins University Ionosphere database

2. Source Information:

   -- Donor: Vince Sigillito (vgs@aplcen.apl.jhu.edu)

   -- Date: 1989

   -- Source: Space Physics Group

              Applied Physics Laboratory

              Johns Hopkins University

              Johns Hopkins Road

              Laurel, MD 20723 

3. Past Usage:

   -- Sigillito, V. G., Wing, S. P., Hutton, L. V., \& Baker, K. B. (1989).

      Classification of radar returns from the ionosphere using neural 

      networks. Johns Hopkins APL Technical Digest, 10, 262-266.

      They investigated using backprop and the perceptron training algorithm

      on this database.  Using the first 200 instances for training, which

      were carefully split almost 50% positive and 50% negative, they found

      that a "linear" perceptron attained 90.7%, a "non-linear" perceptron

      attained 92%, and backprop an average of over 96% accuracy on the 

      remaining 150 test instances, consisting of 123 "good" and only 24 "bad"

      instances.  (There was a counting error or some mistake somewhere; there

      are a total of 351 rather than 350 instances in this domain.) Accuracy

      on "good" instances was much higher than for "bad" instances.  Backprop

      was tested with several different numbers of hidden units (in [0,15])

      and incremental results were also reported (corresponding to how well

      the different variants of backprop did after a periodic number of 

      epochs).

      David Aha (aha@ics.uci.edu) briefly investigated this database.

      He found that nearest neighbor attains an accuracy of 92.1%, that

      Ross Quinlan's C4 algorithm attains 94.0% (no windowing), and that

      IB3 (Aha \& Kibler, IJCAI-1989) attained 96.7% (parameter settings:

      70% and 80% for acceptance and dropping respectively).

4. Relevant Information:

   This radar data was collected by a system in Goose Bay, Labrador.  This

   system consists of a phased array of 16 high-frequency antennas with a

   total transmitted power on the order of 6.4 kilowatts.  See the paper

   for more details.  The targets were free electrons in the ionosphere.

   "Good" radar returns are those showing evidence of some type of structure 

   in the ionosphere.  "Bad" returns are those that do not; their signals pass

   through the ionosphere.  

   Received signals were processed using an autocorrelation function whose

   arguments are the time of a pulse and the pulse number.  There were 17

   pulse numbers for the Goose Bay system.  Instances in this databse are

   described by 2 attributes per pulse number, corresponding to the complex

   values returned by the function resulting from the complex electromagnetic

   signal.

5. Number of Instances: 351

6. Number of Attributes: 34 plus the class attribute

   -- All 34 predictor attributes are continuous

7. Attribute Information:     

   -- All 34 are continuous, as described above

   -- The 35th attribute is either "good" or "bad" according to the definition

      summarized above.  This is a binary classification task.

8. 缺失值：无