def test_arc_reversals(self, configs_cache, mut_inf_cache, return_queue):
print("Test Reversals")
### TEST ARC REVERSALS ###
max_delta = 0
max_operation = None
max_arc = None
max_qi = 0
for u in self.bn.nodes():
for v in self.c_dict[u]:
if not would_cause_cycle(self.c_dict, v, u, reverse=True) and (
self.restriction is None or (v, u) in self.restriction): # and (
# self.whitelist is None or (u,v) not in self.whitelist):
# SCORE FOR 'U' -> gaining 'v' as parent
old_cols = (u,) + tuple(self.p_dict[u]) # without 'v' as parent
if old_cols not in mut_inf_cache:
mut_inf_cache[old_cols] = mutual_information(self.data[list(old_cols)])
mi_old = mut_inf_cache[old_cols]
new_cols = old_cols + (v,) # with 'v' as parent
if new_cols not in mut_inf_cache:
mut_inf_cache[new_cols] = mutual_information(self.data[list(new_cols)])
mi_new = mut_inf_cache[new_cols]
delta1 = self.nrow * (mi_new - mi_old) # Add difference in complexity -> recalculate qi for node v
# SCORE FOR 'V' -> losing 'u' as parent
old_cols = (v,) + tuple(self.p_dict[v]) # with 'u' as parent
if old_cols not in mut_inf_cache:
mut_inf_cache[old_cols] = mutual_information(self.data[list(old_cols)])
mi_old = mut_inf_cache[old_cols]
new_cols = tuple([i for i in old_cols if i != u]) # without 'u' as parent
if new_cols not in mut_inf_cache:
mut_inf_cache[new_cols] = mutual_information(self.data[list(new_cols)])
mi_new = mut_inf_cache[new_cols]
delta2 = self.nrow * (mi_new - mi_old) # Add difference in complexity -> recalculate qi for node v
# COMBINED DELTA-SCORES
ri1 = self.bn.F[u]['ri']
qi1 = self.bn.F[u]['qi']
qi_new1 = calc_num_parent_configs(self.data, self.bn.parents(u) + [v], configs_cache)
ri2 = self.bn.F[v]['ri']
qi2 = self.bn.F[v]['qi']
qi_new2 = calc_num_parent_configs(self.data, [x for x in self.bn.parents(v) if x != u],
configs_cache)
delta_score = delta1 + delta2 - (ri2 * (qi_new2 - qi2) - (qi_new2 - qi2)) - (
ri1 * (qi_new1 - qi1) - (
qi_new1 - qi1)) # Add difference in complexity -> recalculate qi for node u and v
if delta_score - max_delta > 10 ** (-10):
max_delta = delta_score
max_operation = 'Reversal'
max_arc = (u, v)
max_qi = (qi_new1, qi_new2)
return_queue.put((max_arc, max_delta, max_operation, max_qi))
def test_arcs(self, configs_cache, l_inf_cache, u, result_queue):
max_delta = 0
max_operation = None
max_arc = None
for v in [
n for n in self.bn.nodes()
if u != n and n not in self.c_dict[u]
and not would_cause_cycle(self.c_dict, u, n)
]:
# FOR MMHC ALGORITHM -> Edge Restrictions
if self.restriction is None or (u, v) in self.restriction:
# SCORE FOR 'V' -> gaining a parent
old_cols = (v, ) + tuple(
self.p_dict[v]) # without 'u' as parent
if len(old_cols) == 1:
if old_cols not in l_inf_cache:
l_inf_cache[old_cols] = calc_score(self.data, old_cols)
l_old = l_inf_cache[old_cols]
else:
if old_cols not in l_inf_cache:
l_inf_cache[old_cols] = calc_score(self.data, old_cols)
if tuple(self.p_dict[v]) not in l_inf_cache:
l_inf_cache[tuple(self.p_dict[v])] = calc_score(
self.data, self.p_dict[v])
l_old = l_inf_cache[old_cols] - l_inf_cache[tuple(
self.p_dict[v])]
new_cols = old_cols + (u, ) # with'u' as parent
if new_cols not in l_inf_cache:
l_inf_cache[new_cols] = calc_score(self.data, new_cols)
new_cols2 = tuple(self.p_dict[v]) + (u, )
if new_cols2 not in l_inf_cache:
l_inf_cache[new_cols2] = calc_score(self.data, new_cols2)
l_new = l_inf_cache[new_cols] - l_inf_cache[new_cols2]
delta_score = (l_new - l_old) - self.sample_size[min(
len(new_cols), len(self.sample_size))]
if delta_score - max_delta > 10**(-10):
max_delta = delta_score
max_operation = 'Addition'
max_arc = (u, v)
result_queue.put((max_arc, max_delta, max_operation))
def test_arcs(self, configs_cache, mut_inf_cache, u, result_queue):
max_delta = 0
max_operation = None
max_arc = None
max_qi = 0
for v in [
n for n in self.bn.nodes()
if u != n and n not in self.c_dict[u]
and not would_cause_cycle(self.c_dict, u, n)
]:
# FOR MMHC ALGORITHM -> Edge Restrictions
if self.restriction is None or (u, v) in self.restriction:
# SCORE FOR 'V' -> gaining a parent
old_cols = (v, ) + tuple(
self.p_dict[v]) # without 'u' as parent
if old_cols not in mut_inf_cache:
mut_inf_cache[old_cols] = mutual_information(
self.data[list(old_cols)])
mi_old = mut_inf_cache[old_cols]
new_cols = old_cols + (u, ) # with'u' as parent
if new_cols not in mut_inf_cache:
mut_inf_cache[new_cols] = mutual_information(
self.data[list(new_cols)])
mi_new = mut_inf_cache[new_cols]
ri = self.bn.F[v]['ri']
qi = self.bn.F[v]['qi']
qi_new = calc_num_parent_configs(self.data,
self.bn.parents(v) + [u],
configs_cache)
delta_score = self.nrow * (mi_new - mi_old) - (
ri * (qi_new - qi) - (qi_new - qi)
) # Add difference in complexity -> recalculate qi for node v
if delta_score - max_delta > 10**(-10):
max_delta = delta_score
max_operation = 'Addition'
max_arc = (u, v)
max_qi = qi_new
result_queue.put((max_arc, max_delta, max_operation, max_qi))
def hc(data, metric='AIC', max_iter=100, debug=False, restriction=None):
"""
Greedy Hill Climbing search proceeds by choosing the move
which maximizes the increase in fitness of the
network at the current step. It continues until
it reaches a point where there does not exist any
feasible single move that increases the network fitness.
It is called "greedy" because it simply does what is
best at the current iteration only, and thus does not
look ahead to what may be better later on in the search.
For computational saving, a Priority Queue (python's heapq)
can be used to maintain the best operators and reduce the
complexity of picking the best operator from O(n^2) to O(nlogn).
This works by maintaining the heapq of operators sorted by their
delta score, and each time a move is made, we only have to recompute
the O(n) delta-scores which were affected by the move. The rest of
the operator delta-scores are not affected.
For additional computational efficiency, we can cache the
sufficient statistics for various families of distributions -
therefore, computing the mutual information for a given family
only needs to happen once.
The possible moves are the following:
- add edge
- delete edge
- invert edge
Arguments
---------
*data* : a nested numpy array
The data from which the Bayesian network
structure will be learned.
*metric* : a string
Which score metric to use.
Options:
- AIC
- BIC / MDL
- LL (log-likelihood)
*max_iter* : an integer
The maximum number of iterations of the
hill-climbing algorithm to run. Note that
the algorithm will terminate on its own if no
improvement is made in a given iteration.
*debug* : boolean
Whether to print(the scores/moves of the)
algorithm as its happening.
*restriction* : a list of 2-tuples
For MMHC algorithm, the list of allowable edge additions.
Returns
-------
*bn* : a BayesNet object
"""
nrow = data.shape[0]
ncol = data.shape[1]
names = range(ncol)
# INITIALIZE NETWORK W/ NO EDGES
# maintain children and parents dict for fast lookups
c_dict = dict([(n, []) for n in names])
p_dict = dict([(n, []) for n in names])
# COMPUTE INITIAL LIKELIHOOD SCORE
value_dict = dict([(n, np.unique(data[:, i])) for i,n in enumerate(names)])
bn = BayesNet(c_dict)
mle_estimator(bn, data)
max_score = info_score(bn, nrow, metric)
# CREATE EMPIRICAL DISTRIBUTION OBJECT FOR CACHING
#ED = EmpiricalDistribution(data,names)
_iter = 0
improvement = True
while improvement:
improvement = False
max_delta = 0
if debug:
print('ITERATION: ', _iter)
### TEST ARC ADDITIONS ###
for u in bn.nodes():
for v in bn.nodes():
if v not in c_dict[u] and u != v and not would_cause_cycle(c_dict, u, v):
# FOR MMHC ALGORITHM -> Edge Restrictions
if restriction is None or (u, v) in restriction:
# SCORE FOR 'V' -> gaining a parent
old_cols = (v,) + tuple(p_dict[v]) # without 'u' as parent
mi_old = mutual_information(data[:, old_cols])
new_cols = old_cols + (u,) # with'u' as parent
mi_new = mutual_information(data[:, new_cols])
delta_score = nrow * (mi_old - mi_new)
if delta_score > max_delta:
#if debug:
# print('Improved Arc Addition: ' , (u,v))
# print('Delta Score: ' , delta_score)
max_delta = delta_score
max_operation = 'Addition'
max_arc = (u,v)
### TEST ARC DELETIONS ###
for u in bn.nodes():
for v in bn.nodes():
if v in c_dict[u]:
# SCORE FOR 'V' -> losing a parent
old_cols = (v,) + tuple(p_dict[v]) # with 'u' as parent
mi_old = mutual_information(data[:,old_cols])
new_cols = tuple([i for i in old_cols if i != u]) # without 'u' as parent
mi_new = mutual_information(data[:,new_cols])
delta_score = nrow * (mi_old - mi_new)
if delta_score > max_delta:
#if debug:
# print('Improved Arc Deletion: ' , (u,v))
# print('Delta Score: ' , delta_score)
max_delta = delta_score
max_operation = 'Deletion'
max_arc = (u,v)
### TEST ARC REVERSALS ###
for u in bn.nodes():
for v in bn.nodes():
if v in c_dict[u] and not would_cause_cycle(c_dict,v,u, reverse=True):
# SCORE FOR 'U' -> gaining 'v' as parent
old_cols = (u,) + tuple(p_dict[v]) # without 'v' as parent
mi_old = mutual_information(data[:,old_cols])
new_cols = old_cols + (v,) # with 'v' as parent
mi_new = mutual_information(data[:,new_cols])
delta1 = nrow * (mi_old - mi_new)
# SCORE FOR 'V' -> losing 'u' as parent
old_cols = (v,) + tuple(p_dict[v]) # with 'u' as parent
mi_old = mutual_information(data[:,old_cols])
new_cols = tuple([u for i in old_cols if i != u]) # without 'u' as parent
mi_new = mutual_information(data[:,new_cols])
delta2 = nrow * (mi_old - mi_new)
# COMBINED DELTA-SCORES
delta_score = delta1 + delta2
if delta_score > max_delta:
#if debug:
# print('Improved Arc Reversal: ' , (u,v))
# print('Delta Score: ' , delta_score)
max_delta = delta_score
max_operation = 'Reversal'
max_arc = (u,v)
### DETERMINE IF/WHERE IMPROVEMENT WAS MADE ###
if max_delta != 0:
improvement = True
u,v = max_arc
if max_operation == 'Addition':
if debug:
print('ADDING: ' , max_arc , '\n')
c_dict[u].append(v)
p_dict[v].append(u)
elif max_operation == 'Deletion':
if debug:
print('DELETING: ' , max_arc , '\n')
c_dict[u].remove(v)
p_dict[v].remove(u)
elif max_operation == 'Reversal':
if debug:
print('REVERSING: ' , max_arc, '\n')
c_dict[u].remove(v)
p_dict[v].remove(u)
c_dict[v].append(u)
p_dict[u].append(v)
else:
if debug:
print('No Improvement on Iter: ' , _iter)
### TEST FOR MAX ITERATION ###
_iter += 1
if _iter > max_iter:
if debug:
print('Max Iteration Reached')
break
# bn = BayesNet(c_dict)
# print("bn is: " + str(bn.E))
return c_dict
def hc_rr(data,
M=5,
R=3,
metric='AIC',
max_iter=100,
debug=False,
restriction=None):
"""
Arguments
---------
*data* : a nested numpy array
The data from which the Bayesian network
structure will be learned.
*metric* : a string
Which score metric to use.
Options:
- AIC
- BIC / MDL
- LL (log-likelihood)
*max_iter* : an integer
The maximum number of iterations of the
hill-climbing algorithm to run. Note that
the algorithm will terminate on its own if no
improvement is made in a given iteration.
*debug* : boolean
Whether to print(the scores/moves of the)
algorithm as its happening.
*restriction* : a list of 2-tuples
For MMHC algorithm, the list of allowable edge additions.
Returns
-------
*bn* : a BayesNet object
"""
nrow = data.shape[0]
ncol = data.shape[1]
names = range(ncol)
# INITIALIZE NETWORK W/ NO EDGES
# maintain children and parents dict for fast lookups
c_dict = dict([(n, []) for n in names])
p_dict = dict([(n, []) for n in names])
# COMPUTE INITIAL LIKELIHOOD SCORE
value_dict = dict([(n, np.unique(data[:, i]))
for i, n in enumerate(names)])
bn = BayesNet(c_dict)
mle_estimator(bn, data)
max_score = info_score(bn, nrow, metric)
_iter = 0
improvement = True
_restarts = 0
while improvement:
improvement = False
max_delta = 0
if debug:
print('ITERATION: ', _iter)
### TEST ARC ADDITIONS ###
for u in bn.nodes():
for v in bn.nodes():
if v not in c_dict[u] and u != v and not would_cause_cycle(
c_dict, u, v):
# FOR MMHC ALGORITHM -> Edge Restrictions
if restriction is None or (u, v) in restriction:
# SCORE FOR 'V' -> gaining a parent
old_cols = (v, ) + tuple(
p_dict[v]) # without 'u' as parent
mi_old = mutual_information(data[:, old_cols])
new_cols = old_cols + (u, ) # with'u' as parent
mi_new = mutual_information(data[:, new_cols])
delta_score = nrow * (mi_old - mi_new)
if delta_score > max_delta:
if debug:
print('Improved Arc Addition: ', (u, v))
print('Delta Score: ', delta_score)
max_delta = delta_score
max_operation = 'Addition'
max_arc = (u, v)
### TEST ARC DELETIONS ###
for u in bn.nodes():
for v in bn.nodes():
if v in c_dict[u]:
# SCORE FOR 'V' -> losing a parent
old_cols = (v, ) + tuple(p_dict[v]) # with 'u' as parent
mi_old = mutual_information(data[:, old_cols])
new_cols = tuple([i for i in old_cols
if i != u]) # without 'u' as parent
mi_new = mutual_information(data[:, new_cols])
delta_score = nrow * (mi_old - mi_new)
if delta_score > max_delta:
if debug:
print('Improved Arc Deletion: ', (u, v))
print('Delta Score: ', delta_score)
max_delta = delta_score
max_operation = 'Deletion'
max_arc = (u, v)
### TEST ARC REVERSALS ###
for u in bn.nodes():
for v in bn.nodes():
if v in c_dict[u] and not would_cause_cycle(
c_dict, v, u, reverse=True):
# SCORE FOR 'U' -> gaining 'v' as parent
old_cols = (u, ) + tuple(
p_dict[v]) # without 'v' as parent
mi_old = mutual_information(data[:, old_cols])
new_cols = old_cols + (v, ) # with 'v' as parent
mi_new = mutual_information(data[:, new_cols])
delta1 = nrow * (mi_old - mi_new)
# SCORE FOR 'V' -> losing 'u' as parent
old_cols = (v, ) + tuple(p_dict[v]) # with 'u' as parent
mi_old = mutual_information(data[:, old_cols])
new_cols = tuple([u for i in old_cols
if i != u]) # without 'u' as parent
mi_new = mutual_information(data[:, new_cols])
delta2 = nrow * (mi_old - mi_new)
# COMBINED DELTA-SCORES
delta_score = delta1 + delta2
if delta_score > max_delta:
if debug:
print('Improved Arc Reversal: ', (u, v))
print('Delta Score: ', delta_score)
max_delta = delta_score
max_operation = 'Reversal'
max_arc = (u, v)
### DETERMINE IF/WHERE IMPROVEMENT WAS MADE ###
if max_delta != 0:
improvement = True
u, v = max_arc
if max_operation == 'Addition':
if debug:
print('ADDING: ', max_arc, '\n')
c_dict[u].append(v)
p_dict[v].append(u)
elif max_operation == 'Deletion':
if debug:
print('DELETING: ', max_arc, '\n')
c_dict[u].remove(v)
p_dict[v].remove(u)
elif max_operation == 'Reversal':
if debug:
print('REVERSING: ', max_arc, '\n')
c_dict[u].remove(v)
p_dict[v].remove(u)
c_dict[v].append(u)
p_dict[u].append(v)
else:
if debug:
print('No Improvement on Iter: ', _iter)
#### RESTART WITH RANDOM MOVES ####
if _restarts < R:
improvement = True # make another pass of hill climbing
_iter = 0 # reset iterations
if debug:
print('Restart - ', _restarts)
_restarts += 1
for _ in range(M):
# 0 = Addition, 1 = Deletion, 2 = Reversal
operation = np.random.choice([0, 1, 2])
if operation == 0:
while True:
u, v = np.random.choice(list(bn.nodes()),
size=2,
replace=False)
# IF EDGE DOESN'T EXIST, ADD IT
if u not in p_dict[
v] and u != v and not would_cause_cycle(
c_dict, u, v):
if debug:
print('RESTART - ADDING: ', (u, v))
c_dict[u].append(v)
p_dict[v].append(u)
break
elif operation == 1:
while True:
u, v = np.random.choice(list(bn.nodes()),
size=2,
replace=False)
# IF EDGE EXISTS, DELETE IT
if u in p_dict[v]:
if debug:
print('RESTART - DELETING: ', (u, v))
c_dict[u].remove(v)
p_dict[v].remove(u)
break
elif operation == 2:
while True:
u, v = np.random.choice(list(bn.nodes()),
size=2,
replace=False)
# IF EDGE EXISTS, REVERSE IT
if u in p_dict[v] and not would_cause_cycle(
c_dict, v, u, reverse=True):
if debug:
print('RESTART - REVERSING: ', (u, v))
c_dict[u].remove(v)
p_dict[v].remove(u)
c_dict[v].append(u)
p_dict[u].append(v)
break
### TEST FOR MAX ITERATION ###
_iter += 1
if _iter > max_iter:
if debug:
print('Max Iteration Reached')
break
# bn = BayesNet(c_dict)
return c_dict
def tabu(data, k=5, metric='AIC', max_iter=100, debug=False, restriction=None):
"""
Tabu search for score-based structure learning.
The algorithm maintains a list called "tabu_list",
which consists of 3-tuples, where the first two
elements constitute the edge which is tabued, and
the third element is a string - either 'Addition',
'Deletion', or 'Reversal' denoting the operation
associated with the edge.
Arguments
---------
*data* : a nested numpy array
The data from which the Bayesian network
structure will be learned.
*metric* : a string
Which score metric to use.
Options:
- AIC
- BIC / MDL
- LL (log-likelihood)
*max_iter* : an integer
The maximum number of iterations of the
hill-climbing algorithm to run. Note that
the algorithm will terminate on its own if no
improvement is made in a given iteration.
*debug* : boolean
Whether to print(the scores/moves of the)
algorithm as its happening.
*restriction* : a list of 2-tuples
For MMHC algorithm, the list of allowable edge additions.
Returns
-------
*bn* : a BayesNet object
"""
nrow = data.shape[0]
ncol = data.shape[1]
names = range(ncol)
# INITIALIZE NETWORK W/ NO EDGES
# maintain children and parents dict for fast lookups
c_dict = dict([(n, []) for n in names])
p_dict = dict([(n, []) for n in names])
# COMPUTE INITIAL LIKELIHOOD SCORE
value_dict = dict([(n, np.unique(data[:, i]))
for i, n in enumerate(names)])
bn = BayesNet(c_dict)
mle_estimator(bn, data)
max_score = info_score(bn, nrow, metric)
tabu_list = [None] * k
_iter = 0
improvement = True
while improvement:
improvement = False
max_delta = 0
if debug:
print('ITERATION: ', _iter)
### TEST ARC ADDITIONS ###
for u in bn.nodes():
for v in bn.nodes():
# CHECK TABU LIST - can't delete an addition on the tabu list
if (u, v, 'Deletion') not in tabu_list:
# CHECK EDGE EXISTENCE AND CYCLICITY
if v not in c_dict[u] and u != v and not would_cause_cycle(
c_dict, u, v):
# FOR MMHC ALGORITHM -> Edge Restrictions
if restriction is None or (u, v) in restriction:
# SCORE FOR 'V' -> gaining a parent
old_cols = (v, ) + tuple(
p_dict[v]) # without 'u' as parent
mi_old = mutual_information(data[:, old_cols])
new_cols = old_cols + (u, ) # with'u' as parent
mi_new = mutual_information(data[:, new_cols])
delta_score = nrow * (mi_old - mi_new)
if delta_score > max_delta:
if debug:
print('Improved Arc Addition: ', (u, v))
print('Delta Score: ', delta_score)
max_delta = delta_score
max_operation = 'Addition'
max_arc = (u, v)
### TEST ARC DELETIONS ###
for u in bn.nodes():
for v in bn.nodes():
# CHECK TABU LIST - can't add back a deletion on the tabu list
if (u, v, 'Addition') not in tabu_list:
if v in c_dict[u]:
# SCORE FOR 'V' -> losing a parent
old_cols = (v, ) + tuple(
p_dict[v]) # with 'u' as parent
mi_old = mutual_information(data[:, old_cols])
new_cols = tuple([i for i in old_cols
if i != u]) # without 'u' as parent
mi_new = mutual_information(data[:, new_cols])
delta_score = nrow * (mi_old - mi_new)
if delta_score > max_delta:
if debug:
print('Improved Arc Deletion: ', (u, v))
print('Delta Score: ', delta_score)
max_delta = delta_score
max_operation = 'Deletion'
max_arc = (u, v)
### TEST ARC REVERSALS ###
for u in bn.nodes():
for v in bn.nodes():
# CHECK TABU LIST - can't reverse back a reversal on the tabu list
if (u, v, 'Reversal') not in tabu_list:
if v in c_dict[u] and not would_cause_cycle(
c_dict, v, u, reverse=True):
# SCORE FOR 'U' -> gaining 'v' as parent
old_cols = (u, ) + tuple(
p_dict[v]) # without 'v' as parent
mi_old = mutual_information(data[:, old_cols])
new_cols = old_cols + (v, ) # with 'v' as parent
mi_new = mutual_information(data[:, new_cols])
delta1 = nrow * (mi_old - mi_new)
# SCORE FOR 'V' -> losing 'u' as parent
old_cols = (v, ) + tuple(
p_dict[v]) # with 'u' as parent
mi_old = mutual_information(data[:, old_cols])
new_cols = tuple([u for i in old_cols
if i != u]) # without 'u' as parent
mi_new = mutual_information(data[:, new_cols])
delta2 = nrow * (mi_old - mi_new)
# COMBINED DELTA-SCORES
delta_score = delta1 + delta2
if delta_score > max_delta:
if debug:
print('Improved Arc Reversal: ', (u, v))
print('Delta Score: ', delta_score)
max_delta = delta_score
max_operation = 'Reversal'
max_arc = (u, v)
### DETERMINE IF/WHERE IMPROVEMENT WAS MADE ###
if max_delta != 0:
improvement = True
u, v = max_arc
if max_operation == 'Addition':
if debug:
print('ADDING: ', max_arc, '\n')
c_dict[u].append(v)
p_dict[v].append(u)
tabu_list[_iter % 5] = (u, v, 'Addition')
elif max_operation == 'Deletion':
if debug:
print('DELETING: ', max_arc, '\n')
c_dict[u].remove(v)
p_dict[v].remove(u)
tabu_list[_iter % 5] = (u, v, 'Deletion')
elif max_operation == 'Reversal':
if debug:
print('REVERSING: ', max_arc, '\n')
c_dict[u].remove(v)
p_dict[v].remove(u)
c_dict[v].append(u)
p_dict[u].append(v)
tabu_list[_iter % 5] = (u, v, 'Reversal')
else:
if debug:
print('No Improvement on Iter: ', _iter)
### TEST FOR MAX ITERATION ###
_iter += 1
if _iter > max_iter:
if debug:
print('Max Iteration Reached')
break
# bn = BayesNet(c_dict)
return bn
def hc_rr(data, M=5, R=3, metric='AIC', max_iter=100, debug=False, restriction=None): """ Arguments --------- *data* : a nested numpy array The data from which the Bayesian network structure will be learned. *metric* : a string Which score metric to use. Options: - AIC - BIC / MDL - LL (log-likelihood) *max_iter* : an integer The maximum number of iterations of the hill-climbing algorithm to run. Note that the algorithm will terminate on its own if no improvement is made in a given iteration. *debug* : boolean Whether to print the scores/moves of the algorithm as its happening. *restriction* : a list of 2-tuples For MMHC algorithm, the list of allowable edge additions. Returns ------- *bn* : a BayesNet object """ nrow = data.shape[0] ncol = data.shape[1] names = range(ncol) # INITIALIZE NETWORK W/ NO EDGES # maintain children and parents dict for fast lookups c_dict = dict([(n,[]) for n in names]) p_dict = dict([(n,[]) for n in names]) # COMPUTE INITIAL LIKELIHOOD SCORE value_dict = dict([(n, np.unique(data[:,i])) for i,n in enumerate(names)]) bn = BayesNet(c_dict) mle_estimator(bn, data) max_score = info_score(bn, nrow, metric) _iter = 0 improvement = True _restarts = 0 while improvement: improvement = False max_delta = 0 if debug: print 'ITERATION: ' , _iter ### TEST ARC ADDITIONS ### for u in bn.nodes(): for v in bn.nodes(): if v not in c_dict[u] and u!=v and not would_cause_cycle(c_dict, u, v): # FOR MMHC ALGORITHM -> Edge Restrictions if restriction is None or (u,v) in restriction: # SCORE FOR 'V' -> gaining a parent old_cols = (v,) + tuple(p_dict[v]) # without 'u' as parent mi_old = mutual_information(data[:,old_cols]) new_cols = old_cols + (u,) # with'u' as parent mi_new = mutual_information(data[:,new_cols]) delta_score = nrow * (mi_old - mi_new) if delta_score > max_delta: if debug: print 'Improved Arc Addition: ' , (u,v) print 'Delta Score: ' , delta_score max_delta = delta_score max_operation = 'Addition' max_arc = (u,v) ### TEST ARC DELETIONS ### for u in bn.nodes(): for v in bn.nodes(): if v in c_dict[u]: # SCORE FOR 'V' -> losing a parent old_cols = (v,) + tuple(p_dict[v]) # with 'u' as parent mi_old = mutual_information(data[:,old_cols]) new_cols = tuple([i for i in old_cols if i != u]) # without 'u' as parent mi_new = mutual_information(data[:,new_cols]) delta_score = nrow * (mi_old - mi_new) if delta_score > max_delta: if debug: print 'Improved Arc Deletion: ' , (u,v) print 'Delta Score: ' , delta_score max_delta = delta_score max_operation = 'Deletion' max_arc = (u,v) ### TEST ARC REVERSALS ### for u in bn.nodes(): for v in bn.nodes(): if v in c_dict[u] and not would_cause_cycle(c_dict,v,u, reverse=True): # SCORE FOR 'U' -> gaining 'v' as parent old_cols = (u,) + tuple(p_dict[v]) # without 'v' as parent mi_old = mutual_information(data[:,old_cols]) new_cols = old_cols + (v,) # with 'v' as parent mi_new = mutual_information(data[:,new_cols]) delta1 = nrow * (mi_old - mi_new) # SCORE FOR 'V' -> losing 'u' as parent old_cols = (v,) + tuple(p_dict[v]) # with 'u' as parent mi_old = mutual_information(data[:,old_cols]) new_cols = tuple([u for i in old_cols if i != u]) # without 'u' as parent mi_new = mutual_information(data[:,new_cols]) delta2 = nrow * (mi_old - mi_new) # COMBINED DELTA-SCORES delta_score = delta1 + delta2 if delta_score > max_delta: if debug: print 'Improved Arc Reversal: ' , (u,v) print 'Delta Score: ' , delta_score max_delta = delta_score max_operation = 'Reversal' max_arc = (u,v) ### DETERMINE IF/WHERE IMPROVEMENT WAS MADE ### if max_delta != 0: improvement = True u,v = max_arc if max_operation == 'Addition': if debug: print 'ADDING: ' , max_arc , '\n' c_dict[u].append(v) p_dict[v].append(u) elif max_operation == 'Deletion': if debug: print 'DELETING: ' , max_arc , '\n' c_dict[u].remove(v) p_dict[v].remove(u) elif max_operation == 'Reversal': if debug: print 'REVERSING: ' , max_arc, '\n' c_dict[u].remove(v) p_dict[v].remove(u) c_dict[v].append(u) p_dict[u].append(v) else: if debug: print 'No Improvement on Iter: ' , _iter #### RESTART WITH RANDOM MOVES #### if _restarts < R: improvement = True # make another pass of hill climbing _iter=0 # reset iterations if debug: print 'Restart - ' , _restarts _restarts+=1 for _ in range(M): # 0 = Addition, 1 = Deletion, 2 = Reversal operation = np.random.choice([0,1,2]) if operation == 0: while True: u,v = np.random.choice(list(bn.nodes()), size=2, replace=False) # IF EDGE DOESN'T EXIST, ADD IT if u not in p_dict[v] and u!=v and not would_cause_cycle(c_dict,u,v): if debug: print 'RESTART - ADDING: ', (u,v) c_dict[u].append(v) p_dict[v].append(u) break elif operation == 1: while True: u,v = np.random.choice(list(bn.nodes()), size=2, replace=False) # IF EDGE EXISTS, DELETE IT if u in p_dict[v]: if debug: print 'RESTART - DELETING: ', (u,v) c_dict[u].remove(v) p_dict[v].remove(u) break elif operation == 2: while True: u,v = np.random.choice(list(bn.nodes()), size=2, replace=False) # IF EDGE EXISTS, REVERSE IT if u in p_dict[v] and not would_cause_cycle(c_dict,v,u, reverse=True): if debug: print 'RESTART - REVERSING: ', (u,v) c_dict[u].remove(v) p_dict[v].remove(u) c_dict[v].append(u) p_dict[u].append(v) break ### TEST FOR MAX ITERATION ### _iter += 1 if _iter > max_iter: if debug: print 'Max Iteration Reached' break bn = BayesNet(c_dict) return bn
def hc(data, metric='AIC', max_iter=100, debug=False, restriction=None): """ Greedy Hill Climbing search proceeds by choosing the move which maximizes the increase in fitness of the network at the current step. It continues until it reaches a point where there does not exist any feasible single move that increases the network fitness. It is called "greedy" because it simply does what is best at the current iteration only, and thus does not look ahead to what may be better later on in the search. For computational saving, a Priority Queue (python's heapq) can be used to maintain the best operators and reduce the complexity of picking the best operator from O(n^2) to O(nlogn). This works by maintaining the heapq of operators sorted by their delta score, and each time a move is made, we only have to recompute the O(n) delta-scores which were affected by the move. The rest of the operator delta-scores are not affected. For additional computational efficiency, we can cache the sufficient statistics for various families of distributions - therefore, computing the mutual information for a given family only needs to happen once. The possible moves are the following: - add edge - delete edge - invert edge Arguments --------- *data* : a nested numpy array The data from which the Bayesian network structure will be learned. *metric* : a string Which score metric to use. Options: - AIC - BIC / MDL - LL (log-likelihood) *max_iter* : an integer The maximum number of iterations of the hill-climbing algorithm to run. Note that the algorithm will terminate on its own if no improvement is made in a given iteration. *debug* : boolean Whether to print the scores/moves of the algorithm as its happening. *restriction* : a list of 2-tuples For MMHC algorithm, the list of allowable edge additions. Returns ------- *bn* : a BayesNet object """ nrow = data.shape[0] ncol = data.shape[1] names = range(ncol) # INITIALIZE NETWORK W/ NO EDGES # maintain children and parents dict for fast lookups c_dict = dict([(n,[]) for n in names]) p_dict = dict([(n,[]) for n in names]) # COMPUTE INITIAL LIKELIHOOD SCORE value_dict = dict([(n, np.unique(data[:,i])) for i,n in enumerate(names)]) bn = BayesNet(c_dict) mle_estimator(bn, data) max_score = info_score(bn, nrow, metric) # CREATE EMPIRICAL DISTRIBUTION OBJECT FOR CACHING #ED = EmpiricalDistribution(data,names) _iter = 0 improvement = True while improvement: improvement = False max_delta = 0 if debug: print 'ITERATION: ' , _iter ### TEST ARC ADDITIONS ### for u in bn.nodes(): for v in bn.nodes(): if v not in c_dict[u] and u!=v and not would_cause_cycle(c_dict, u, v): # FOR MMHC ALGORITHM -> Edge Restrictions if restriction is None or (u,v) in restriction: # SCORE FOR 'V' -> gaining a parent old_cols = (v,) + tuple(p_dict[v]) # without 'u' as parent mi_old = mutual_information(data[:,old_cols]) new_cols = old_cols + (u,) # with'u' as parent mi_new = mutual_information(data[:,new_cols]) delta_score = nrow * (mi_old - mi_new) if delta_score > max_delta: #if debug: # print 'Improved Arc Addition: ' , (u,v) # print 'Delta Score: ' , delta_score max_delta = delta_score max_operation = 'Addition' max_arc = (u,v) ### TEST ARC DELETIONS ### for u in bn.nodes(): for v in bn.nodes(): if v in c_dict[u]: # SCORE FOR 'V' -> losing a parent old_cols = (v,) + tuple(p_dict[v]) # with 'u' as parent mi_old = mutual_information(data[:,old_cols]) new_cols = tuple([i for i in old_cols if i != u]) # without 'u' as parent mi_new = mutual_information(data[:,new_cols]) delta_score = nrow * (mi_old - mi_new) if delta_score > max_delta: #if debug: # print 'Improved Arc Deletion: ' , (u,v) # print 'Delta Score: ' , delta_score max_delta = delta_score max_operation = 'Deletion' max_arc = (u,v) ### TEST ARC REVERSALS ### for u in bn.nodes(): for v in bn.nodes(): if v in c_dict[u] and not would_cause_cycle(c_dict,v,u, reverse=True): # SCORE FOR 'U' -> gaining 'v' as parent old_cols = (u,) + tuple(p_dict[v]) # without 'v' as parent mi_old = mutual_information(data[:,old_cols]) new_cols = old_cols + (v,) # with 'v' as parent mi_new = mutual_information(data[:,new_cols]) delta1 = nrow * (mi_old - mi_new) # SCORE FOR 'V' -> losing 'u' as parent old_cols = (v,) + tuple(p_dict[v]) # with 'u' as parent mi_old = mutual_information(data[:,old_cols]) new_cols = tuple([u for i in old_cols if i != u]) # without 'u' as parent mi_new = mutual_information(data[:,new_cols]) delta2 = nrow * (mi_old - mi_new) # COMBINED DELTA-SCORES delta_score = delta1 + delta2 if delta_score > max_delta: #if debug: # print 'Improved Arc Reversal: ' , (u,v) # print 'Delta Score: ' , delta_score max_delta = delta_score max_operation = 'Reversal' max_arc = (u,v) ### DETERMINE IF/WHERE IMPROVEMENT WAS MADE ### if max_delta != 0: improvement = True u,v = max_arc if max_operation == 'Addition': if debug: print 'ADDING: ' , max_arc , '\n' c_dict[u].append(v) p_dict[v].append(u) elif max_operation == 'Deletion': if debug: print 'DELETING: ' , max_arc , '\n' c_dict[u].remove(v) p_dict[v].remove(u) elif max_operation == 'Reversal': if debug: print 'REVERSING: ' , max_arc, '\n' c_dict[u].remove(v) p_dict[v].remove(u) c_dict[v].append(u) p_dict[u].append(v) else: if debug: print 'No Improvement on Iter: ' , _iter ### TEST FOR MAX ITERATION ### _iter += 1 if _iter > max_iter: if debug: print 'Max Iteration Reached' break bn = BayesNet(c_dict) return bn
def tabu(data, k=5, metric='AIC', max_iter=100, debug=False, restriction=None): """ Tabu search for score-based structure learning. The algorithm maintains a list called "tabu_list", which consists of 3-tuples, where the first two elements constitute the edge which is tabued, and the third element is a string - either 'Addition', 'Deletion', or 'Reversal' denoting the operation associated with the edge. Arguments --------- *data* : a nested numpy array The data from which the Bayesian network structure will be learned. *metric* : a string Which score metric to use. Options: - AIC - BIC / MDL - LL (log-likelihood) *max_iter* : an integer The maximum number of iterations of the hill-climbing algorithm to run. Note that the algorithm will terminate on its own if no improvement is made in a given iteration. *debug* : boolean Whether to print the scores/moves of the algorithm as its happening. *restriction* : a list of 2-tuples For MMHC algorithm, the list of allowable edge additions. Returns ------- *bn* : a BayesNet object """ nrow = data.shape[0] ncol = data.shape[1] names = range(ncol) # INITIALIZE NETWORK W/ NO EDGES # maintain children and parents dict for fast lookups c_dict = dict([(n,[]) for n in names]) p_dict = dict([(n,[]) for n in names]) # COMPUTE INITIAL LIKELIHOOD SCORE value_dict = dict([(n, np.unique(data[:,i])) for i,n in enumerate(names)]) bn = BayesNet(c_dict) mle_estimator(bn, data) max_score = info_score(bn, nrow, metric) tabu_list = [None]*k _iter = 0 improvement = True while improvement: improvement = False max_delta = 0 if debug: print 'ITERATION: ' , _iter ### TEST ARC ADDITIONS ### for u in bn.nodes(): for v in bn.nodes(): # CHECK TABU LIST - can't delete an addition on the tabu list if (u,v,'Deletion') not in tabu_list: # CHECK EDGE EXISTENCE AND CYCLICITY if v not in c_dict[u] and u!=v and not would_cause_cycle(c_dict, u, v): # FOR MMHC ALGORITHM -> Edge Restrictions if restriction is None or (u,v) in restriction: # SCORE FOR 'V' -> gaining a parent old_cols = (v,) + tuple(p_dict[v]) # without 'u' as parent mi_old = mutual_information(data[:,old_cols]) new_cols = old_cols + (u,) # with'u' as parent mi_new = mutual_information(data[:,new_cols]) delta_score = nrow * (mi_old - mi_new) if delta_score > max_delta: if debug: print 'Improved Arc Addition: ' , (u,v) print 'Delta Score: ' , delta_score max_delta = delta_score max_operation = 'Addition' max_arc = (u,v) ### TEST ARC DELETIONS ### for u in bn.nodes(): for v in bn.nodes(): # CHECK TABU LIST - can't add back a deletion on the tabu list if (u,v,'Addition') not in tabu_list: if v in c_dict[u]: # SCORE FOR 'V' -> losing a parent old_cols = (v,) + tuple(p_dict[v]) # with 'u' as parent mi_old = mutual_information(data[:,old_cols]) new_cols = tuple([i for i in old_cols if i != u]) # without 'u' as parent mi_new = mutual_information(data[:,new_cols]) delta_score = nrow * (mi_old - mi_new) if delta_score > max_delta: if debug: print 'Improved Arc Deletion: ' , (u,v) print 'Delta Score: ' , delta_score max_delta = delta_score max_operation = 'Deletion' max_arc = (u,v) ### TEST ARC REVERSALS ### for u in bn.nodes(): for v in bn.nodes(): # CHECK TABU LIST - can't reverse back a reversal on the tabu list if (u,v,'Reversal') not in tabu_list: if v in c_dict[u] and not would_cause_cycle(c_dict,v,u, reverse=True): # SCORE FOR 'U' -> gaining 'v' as parent old_cols = (u,) + tuple(p_dict[v]) # without 'v' as parent mi_old = mutual_information(data[:,old_cols]) new_cols = old_cols + (v,) # with 'v' as parent mi_new = mutual_information(data[:,new_cols]) delta1 = nrow * (mi_old - mi_new) # SCORE FOR 'V' -> losing 'u' as parent old_cols = (v,) + tuple(p_dict[v]) # with 'u' as parent mi_old = mutual_information(data[:,old_cols]) new_cols = tuple([u for i in old_cols if i != u]) # without 'u' as parent mi_new = mutual_information(data[:,new_cols]) delta2 = nrow * (mi_old - mi_new) # COMBINED DELTA-SCORES delta_score = delta1 + delta2 if delta_score > max_delta: if debug: print 'Improved Arc Reversal: ' , (u,v) print 'Delta Score: ' , delta_score max_delta = delta_score max_operation = 'Reversal' max_arc = (u,v) ### DETERMINE IF/WHERE IMPROVEMENT WAS MADE ### if max_delta != 0: improvement = True u,v = max_arc if max_operation == 'Addition': if debug: print 'ADDING: ' , max_arc , '\n' c_dict[u].append(v) p_dict[v].append(u) tabu_list[_iter % 5] = (u,v,'Addition') elif max_operation == 'Deletion': if debug: print 'DELETING: ' , max_arc , '\n' c_dict[u].remove(v) p_dict[v].remove(u) tabu_list[_iter % 5] = (u,v,'Deletion') elif max_operation == 'Reversal': if debug: print 'REVERSING: ' , max_arc, '\n' c_dict[u].remove(v) p_dict[v].remove(u) c_dict[v].append(u) p_dict[u].append(v) tabu_list[_iter % 5] = (u,v,'Reversal') else: if debug: print 'No Improvement on Iter: ' , _iter ### TEST FOR MAX ITERATION ### _iter += 1 if _iter > max_iter: if debug: print 'Max Iteration Reached' break bn = BayesNet(c_dict) return bn
