SHOW:
|
|
- or go back to the newest paste.
| 1 | from django.core.management.base import BaseCommand, CommandError | |
| 2 | from math import sqrt, fabs, pow | |
| 3 | from copy import deepcopy | |
| 4 | from collections import defaultdict | |
| 5 | ||
| 6 | from ratings.models import Rating, Participant, ParticipantRelation | |
| 7 | ||
| 8 | class RelationEqualizer(object): | |
| 9 | def __init__(self, rating): | |
| 10 | self._rating = rating | |
| 11 | self._prel = {}
| |
| 12 | self._pmap = {}
| |
| 13 | self._antifraud = {}
| |
| 14 | self._uvn_iterations = 10 # user vector normalization iterations | |
| 15 | self._tn_iterations = 5 # table normalization iterations | |
| 16 | ||
| 17 | def _print(self): | |
| 18 | skeys = sorted(self._pmap.keys(), key=lambda x: -self._pmap[x].score) | |
| 19 | for row_id in skeys: | |
| 20 | vals = [ u'%20s' % self._pmap[row_id].user.username ] | |
| 21 | total = 0.0 | |
| 22 | count = 0.0 | |
| 23 | for col_id in skeys: | |
| 24 | rel = self._prel[row_id][col_id] | |
| 25 | if rel is None: | |
| 26 | vals.append('U')
| |
| 27 | elif rel > 0: | |
| 28 | vals.append('B')
| |
| 29 | elif rel < 0: | |
| 30 | vals.append('.')
| |
| 31 | else: | |
| 32 | vals.append('=')
| |
| 33 | vals.append('%5.1f' % self._user_scores(row_id))
| |
| 34 | print u' '.join(vals) | |
| 35 | ||
| 36 | def _user_scores(self, user): | |
| 37 | count = len(self._pmap) | |
| 38 | if count == 0: | |
| 39 | return 0.0 | |
| 40 | score = 0.0 | |
| 41 | ||
| 42 | for rel in self._prel[user].itervalues(): | |
| 43 | if rel is None: | |
| 44 | continue | |
| 45 | if rel >= 0: | |
| 46 | score += 1 | |
| 47 | else: | |
| 48 | score -= 1 | |
| 49 | return pow((score / count + 1.0) * 50, 0.5) * 10 | |
| 50 | ||
| 51 | def _load_data(self): | |
| 52 | pmap = {}
| |
| 53 | prel = {}
| |
| 54 | ||
| 55 | rlist = ParticipantRelation.objects.filter(subject__rating=self._rating).all() | |
| 56 | for r in rlist: | |
| 57 | if not r.object.active or not r.subject.active: | |
| 58 | continue | |
| 59 | if not r.subject.pk in pmap: | |
| 60 | pmap[r.subject.pk] = r.subject | |
| 61 | if not r.object.pk in pmap: | |
| 62 | pmap[r.object.pk] = r.object | |
| 63 | if r.subject.pk not in prel: | |
| 64 | prel[r.subject.pk] = {}
| |
| 65 | if r.object.pk not in prel: | |
| 66 | prel[r.object.pk] = {}
| |
| 67 | prel[r.subject.pk][r.object.pk] = r.value | |
| 68 | ||
| 69 | for x in pmap.iterkeys(): | |
| 70 | for y in pmap.iterkeys(): | |
| 71 | if y not in prel[x]: | |
| 72 | prel[x][y] = None | |
| 73 | if x == y: | |
| 74 | prel[x][y] = 0.0 | |
| 75 | ||
| 76 | self._pmap = pmap | |
| 77 | self._prel = prel | |
| 78 | ||
| 79 | def _load_antifraud(self): | |
| 80 | antifraud = {}
| |
| 81 | prel = self._prel | |
| 82 | competitors_number = len(prel) | |
| 83 | for xid, rels in prel.iteritems(): | |
| 84 | antifraud[xid] = 1.0 | |
| 85 | count = 0.0 | |
| 86 | frauds = 0.0 | |
| 87 | for yid, rel in rels.iteritems(): | |
| 88 | if prel[xid][yid] is None or prel[yid][xid] is None: | |
| 89 | continue | |
| 90 | ||
| 91 | count += 1 | |
| 92 | if prel[xid][yid] * prel[yid][xid] < 0 or prel[xid][yid] == 0 and prel[yid][xid] == 0: | |
| 93 | continue | |
| 94 | ||
| 95 | if prel[xid][yid] * prel[yid][xid] > 0: | |
| 96 | frauds += 1.0 | |
| 97 | elif prel[xid][yid] * prel[yid][xid] == 0: | |
| 98 | frauds += 2.0 /3.0 | |
| 99 | if count > 0: | |
| 100 | antifraud[xid] = pow(1.0 - frauds / count, 5) | |
| 101 | self._antifraud = antifraud | |
| 102 | ||
| 103 | def _stronger_weaker_equals(self, id, weight, s, e, w): | |
| 104 | ||
| 105 | for yidx in self._prel.iterkeys(): | |
| 106 | rel = self._prel[id][yidx] | |
| 107 | if rel is not None: | |
| 108 | if rel > 0: | |
| 109 | w[yidx] += weight * self._antifraud[id] | |
| 110 | elif rel < 0: | |
| 111 | s[yidx] += weight * self._antifraud[id] | |
| 112 | elif rel == 0: | |
| 113 | e[yidx] += weight * self._antifraud[id] | |
| 114 | ||
| 115 | rel = self._prel[yidx][id] | |
| 116 | if rel is not None: | |
| 117 | if rel < 0: | |
| 118 | w[yidx] += weight * self._antifraud[yidx] | |
| 119 | elif rel > 0: | |
| 120 | s[yidx] += weight * self._antifraud[yidx] | |
| 121 | elif rel == 0: | |
| 122 | e[yidx] += weight * self._antifraud[yidx] | |
| 123 | ||
| 124 | def _update_user_vector(self, user): | |
| 125 | ||
| 126 | stronger = defaultdict(lambda: 0.0) # players that stronger than user | |
| 127 | equals = defaultdict(lambda: 0.0) | |
| 128 | weaker = defaultdict(lambda: 0.0) | |
| 129 | equals[user] = 1 | |
| 130 | ||
| 131 | for i in xrange(0, self._uvn_iterations): | |
| 132 | dummy = defaultdict(lambda: 0.0) | |
| 133 | s = defaultdict(lambda: 0.0) | |
| 134 | w = defaultdict(lambda: 0.0) | |
| 135 | e = defaultdict(lambda: 0.0) | |
| 136 | ||
| 137 | for yidx, yval in stronger.iteritems(): | |
| 138 | self._stronger_weaker_equals(yidx, yval, s, s, dummy) | |
| 139 | ||
| 140 | for yidx, yval in equals.iteritems(): | |
| 141 | self._stronger_weaker_equals(yidx, yval, s, e, w) | |
| 142 | ||
| 143 | for yidx, yval in weaker.iteritems(): | |
| 144 | self._stronger_weaker_equals(yidx, yval, dummy, w, w) | |
| 145 | ||
| 146 | for k in set(s.keys()) | set(e.keys()) | set(w.keys()): | |
| 147 | sum = s[k] + w[k] + e[k] | |
| 148 | if sum > 0: | |
| 149 | s[k] /= sum | |
| 150 | e[k] /= sum | |
| 151 | w[k] /= sum | |
| 152 | ||
| 153 | stronger = s | |
| 154 | equals = e | |
| 155 | weaker = w | |
| 156 | ||
| 157 | result = {}
| |
| 158 | for idx in self._prel.iterkeys(): | |
| 159 | result[idx] = w[idx] - s[idx] | |
| 160 | if fabs(result[idx]) <= 1.0 / 3: | |
| 161 | result[idx] = 0.0 | |
| 162 | return result | |
| 163 | ||
| 164 | def relations(self): | |
| 165 | self._load_data() | |
| 166 | for i in xrange(0, self._tn_iterations): | |
| 167 | new_prel = {}
| |
| 168 | self._load_antifraud() | |
| 169 | for user in self._prel.iterkeys(): | |
| 170 | new_prel[user] = self._update_user_vector(user) | |
| 171 | self._prel = new_prel | |
| 172 | return self._prel | |
| 173 | ||
| 174 | def update_rates(self): | |
| 175 | self._load_data() | |
| 176 | self._print() | |
| 177 | prel = self.relations() | |
| 178 | print "" | |
| 179 | self._print() | |
| 180 | for user in prel.iterkeys(): | |
| 181 | uname = self._pmap[user].user.username.encode('utf8', 'ignore')
| |
| 182 | self._pmap[user].score = self._user_scores(user) | |
| 183 | self._pmap[user].save() | |
| 184 | print "User %s scores %f" % (uname, self._pmap[user].score) | |
| 185 | ||
| 186 | class Command(BaseCommand): | |
| 187 | help = 'ratings normalization' | |
| 188 | ||
| 189 | def handle(self, *args, **kwargs): | |
| 190 | for r in Rating.objects.all(): | |
| 191 | self.stdout.write("Processing rating %s" % r.title)
| |
| 192 | req = RelationEqualizer(r) | |
| 193 | req.update_rates() | |
| 194 | self.stdout.write("") |
