def measure(self, target: Labelling, predicted) -> float:
avg_prec = 0.0
npos = target.positives()
if (npos == 0):
return 1.0
pos_idxs = target.where_positive()
for i in pos_idxs:
yi_rank = predicted[i]
s = sum([1 for j in pos_idxs if predicted[j] <= yi_rank])
avg_prec += s / yi_rank
return avg_prec / npos
def clr(P: ProbabilityDistribution) -> Labelling:
n = P.getNumberOfLabels()
Pm = P.marginal()
Pmsum = n - sum(Pm)
return Labelling([
1 if (_rpc_labelscore(P, i) + Pm[i] >= Pmsum) else 0 for i in range(n)
])
def minrisk(self, Pdist: ProbabilityDistribution):
min_risk = 128.0
for L in Labelling.iterAll(Pdist.getNumberOfLabels()):
r = self.risk(L, Pdist)
if (r < min_risk):
min_risk = r
min_risk_sol = L
return (min_risk_sol, min_risk)
def maxrisk(self, Pdist: ProbabilityDistribution):
max_risk = 0.0
for L in Labelling.iterAll(Pdist.getNumberOfLabels()):
r = self.risk(L, Pdist)
if (r > max_risk):
max_risk = r
max_risk_sol = L
return (max_risk_sol, max_risk)
def risk(self, prediction: Labelling,
Pdist: ProbabilityDistribution) -> float:
n = Pdist.getNumberOfLabels()
k = prediction.positives()
if (k == 0):
return 1 - Pdist.fulljoint([0] * n)
p = 0.0
for i in range(n):
if (prediction[i] == 1):
p += sum([
Pdist.getProbSize(i, j) / (k + j) for j in range(1, n + 1)
])
return 1 - 2 * p
def dbr(P: ProbabilityDistribution) -> Labelling:
n = P.getNumberOfLabels()
M = P.marginal()
ybr = np.where(M >= 1 / 2, 1, 0)
L = Labelling.zeros(n)
for i in range(n):
vi = ybr[i]
ybr[i] = 0
p0 = P.fulljoint(ybr)
ybr[i] = 1
p1 = P.fulljoint(ybr)
ybr[i] = vi
L[i] = 1 if (p1 + 1e-20) / (p0 + p1 + 2e-20) >= 1 / 2 else 0
return L
def classifier_chain(P: ProbabilityDistribution):
n = P.getNumberOfLabels()
given = np.empty((n, 2), dtype=int)
given[:, 0] = np.arange(n)
ask = np.array([[0, 1]])
perm = np.random.permutation(range(n))
for i, li in enumerate(perm):
ask[0, 0] = li
p = P.jointCond(ask, given[perm[:i]])
if (p > 0.5):
given[li, 1] = 1
else:
given[li, 1] = 0
return Labelling(given[:, 1])
def measure(self, target: Labelling, predicted) -> float:
s = 0
for ri, yi in zip(predicted, target):
if (yi == 0):
continue
for rj, yj in zip(predicted, target):
if (yj == 1):
continue
if (ri > rj):
s += 1
num_pos = target.positives()
n = len(target)
if (self.normalized == 2):
return s / (num_pos * (n - num_pos))
if (self.normalized == 1):
return 4 * s / (n * n)
return s
def minrisk(self, Pdist: ProbabilityDistribution):
n = Pdist.getNumberOfLabels()
def predict_k(k):
D = []
R = range(1, n + 1)
for i in range(n):
s1 = sum([Pdist.getProbSize(i, s) / (s + k) for s in R])
D.append(s1)
pos = np.argsort(D)[-k:]
Y = [0] * n
v = 0.0
for p in pos:
Y[p] = 1
v += 2 * D[p]
return (Y, v)
bests_Y = [([0] * n, Pdist.fulljoint([0] * n))]
bests_Y += [predict_k(i) for i in range(1, n + 1)]
V = [Y[1] for Y in bests_Y]
v = np.argmax(V)
best = bests_Y[v]
# assert(np.isclose(1-best[1], super().minrisk(Pdist)[1]))
return Labelling(best[0]), 1 - best[1]
def br(P) -> Labelling:
M = P.marginal()
return Labelling(np.where(M >= 1 / 2, 1, 0))
def minrisk(self, Pdist: ProbabilityDistribution):
L = Labelling([1 if (p >= 1 / 2) else 0 for p in Pdist.marginal()])
return (L, self.risk(L, Pdist))
def measure(self, target: Labelling, predicted) -> float:
positives_idxs = target.where_positive()
return sum([1 / predicted[pi] for pi in positives_idxs]) / sum(
1 / np.arange(1,
target.positives() + 1))
def measure(self, target: Labelling, predicted) -> float:
positives_idxs = target.where_positive()
if (len(positives_idxs) == 0):
return 0.0
return max([predicted[pi] for pi in positives_idxs
]) # / len(target) / target.positives()