knn Perl knn (or first version of OCaml knn which was same)
  Run with parameters 3 1, which means
    Use overlap for weighing the results
    Neighbour multiplied by 1 and rest add to 1

knn-is  Same as above, but TFIDF is updated at every theorem rather than once at the end

knn-lin-is Same as above, but when computing the distance for a neighbour just add the
  weighs and not the weighs squared


geo 1_XX  Compute neighbours like for k-NN, but take them all, and multiply the weight for
  dependencies by XX%

geo r_XX  Like above but include dependencies of dependencies again further multiplied by
  the same factor.

nb = naive bayes
nb_idf = naive bayes with probabilities for each symbol exponentiated with the idf factor

  nb confidence and prior are as follows
    010 = 0.000000001       p=ni/a
    020 = 0.001             p=0.0002
    030 = 0.000000000000001 p=0.0002
    040 = 0.001             p=0.5
    050 = 0.000000000000001 p=ni/a
    060 = 0.000001          p=0.0002
    070 = 0.00001           0.001

lsi
  1 = dependencies have the weight 1/(no of deps)
  2 = dependencies have the weight 1/2
  3 = dependencies have the weight 1/20
  4 = dependencies have the weight 1/100
  5 = max rather than sum,  dependencies have weight 1/2
  6 = neighbours = exp, deps = 1
  7 = neighbours = exp, deps = square
  8 = neighbours = exp square, deps = square

combine

  Merge sequences with scaled votes, first is NB factor, second is LSI factor, rest is KNN.
  (Or if there is one more factor, then the 3rd factor is for Sine).

  min = minimum vote
  avg = average vote = arithmetic mean (or sum of votes)
  geo = geometric-average of votes (or sum of logs or multiply of votes)
  har = harmonic-mean
  qua = quadratic mean
  max = max


  by default all votes till end are considered
  _1k_ = votes bigger than 1024 are same as voting 1024
  _3k_ = votes bigger than 3072 are considered same as voting 3072
  _10k_ = votes bigger then 10240 are same as voting 10240

linreg: size of B:
  55727x55727
  75369x75369
  31586x56444