def slow_ris3(dataset, targets, threshold, n_jobs=1):
scores, radius = _scores_radius(dataset, targets, n_jobs)
_norm_by_clas(scores, targets)
# Remove from idx the ones that have bad scores and recompute radius
idx = scores.argsort(None)[::-1]
for i, v in enumerate(scores[idx]):
if v < threshold:
idx = idx[:i]
break
# Now recompute radius
for id in idx:
instance = dataset[id, :]
target = targets[id]
max_dist = inf
for o_id in idx:
o_instance = dataset[o_id, :]
o_target = targets[o_id]
if o_target != target:
dist = euclidean(instance, o_instance)
if dist < max_dist:
max_dist = dist
radius[id] = max_dist
return slow_relevants(idx, scores, radius, threshold, dataset, targets)
def slow_relevants(idx, scores, radius, threshold, dataset, targets):
selected_set = []
selected_radius = []
for id in idx:
score = scores[id]
if score < threshold:
break
instance = dataset[id, :]
target = targets[id]
for s_id in selected_set:
another_target = targets[s_id]
if another_target == target:
dist = euclidean(instance, dataset[s_id, :])
if dist < radius[s_id]:
break
else:
selected_set.append(id)
selected_radius.append(radius[id])
return selected_set, selected_radius
def _scores_radius(dataset, targets, n_jobs=1):
scores = np.zeros((targets.size, 1), dtype=np.float_)
radius = np.zeros_like(scores)
# Calculate score
for i, (instance, target) in enumerate(zip(dataset, targets)):
denom = 0
num = 0
radius[i] = np.inf
for another_instance, another_target in zip(dataset, targets):
dist = euclidean(instance, another_instance)
val = exp(-dist)
denom += val
if target == another_target:
num += val
else:
num -= val
if dist < radius[i]:
radius[i] = dist
num -= 1
denom -= 1
scores[i] = num / denom
return scores, radius