crispe.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-
from collections import Counter, defaultdict
from copy import copy
import random
import sys
import time

from crisp import map_to_majority, marginals
from entropy import entropy


def regress(X, Y):
    # target Y, feature X
    max_iterations = 10000
    scx = entropy(X)
    len_dom_y = len(set(Y))
    # print scx,
    f = map_to_majority(X, Y)

    supp_x = list(set(X))
    supp_y = list(set(Y))

    pair = zip(X, Y)
    res = [y - f[x] for x, y in pair]
    cur_res_codelen = entropy(res)

    j = 0
    minimized = True
    while j < max_iterations and minimized:
        random.shuffle(supp_x)
        minimized = False

        for x_to_map in supp_x:
            best_res_codelen = sys.float_info.max
            best_cand_y = None

            for cand_y in supp_y:
                if cand_y == f[x_to_map]:
                    continue

                res = [y - f[x] if x != x_to_map else y -
                       cand_y for x, y in pair]
                res_codelen = entropy(res)

                if res_codelen < best_res_codelen:
                    best_res_codelen = res_codelen
                    best_cand_y = cand_y

            if best_res_codelen < cur_res_codelen:
                cur_res_codelen = best_res_codelen
                f[x_to_map] = best_cand_y
                minimized = True
        j += 1
    # print cur_res_codelen
    return scx + cur_res_codelen


def crispe(X, Y):
    sxtoy = regress(X, Y)
    sytox = regress(Y, X)
    return (sxtoy, sytox)


if __name__ == "__main__":
    from test_benchmark import load_pair
    X, Y = load_pair(99)
    print crispe(X, Y)