def run(self):
"""------------------------------------------------------------
High level function that runs the online influence maximization
algorithm for self.rounds times and reports aggregated regret
------------------------------------------------------------"""
seed_set = np.argsort(self.outdegs)[::-1][:self.seed_size].tolist()
for epoch_idx in np.arange(1, self.epochs + 1):
self.get_context()
# Simulates the chosen seed_set's performance in real world
online_spread, tried_cnts, success_cnts = self.simulate_spread(
seed_set)
# Oracle run
real_infs = self.context_influences(self.context_vector)
self.dump_graph(real_infs, ("tim_" + self.graph_file))
oracle = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"),
self.node_cnt, self.edge_cnt, self.seed_size,
bytes("IC", "ascii"))
oracle_set = list(oracle.get_seed_set(self.epsilon))
oracle = None
oracle_spread, _, _ = self.simulate_spread(oracle_set)
self.regret.append(oracle_spread - online_spread)
self.spread.append(online_spread)
def run(self):
"""------------------------------------------------------------
High level function that runs the online influence maximization
algorithm for self.epochs times and reports aggregated regret
------------------------------------------------------------"""
for epoch_idx in np.arange(1, self.epochs+1):
print(epoch_idx)
self.get_context()
context_idx = self.context_classifier(self.context_vector)
explore_bool = binomial(1, self.epsilons[epoch_idx-1])
# Exploration Epoch, set influence estimates accordingly
if(explore_bool):
inf_ests = [(alpha/(alpha+beta)) * (alpha+math.sqrt((alpha*beta)/(alpha+beta+1)))
if(alpha+beta > 0) else (0) for alpha, beta in zip(self.alphas, self.betas)]
# Exploitation Epoch, set influence estimates accordingly
else:
inf_ests = [(alpha/(alpha+beta)) if(alpha+beta > 0)
else (0) for alpha, beta in zip(self.alphas, self.betas)]
self.dump_graph(inf_ests, ("tim_"+self.graph_file))
timgraph = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"), self.node_cnt, self.edge_cnt,
self.seed_size, bytes("IC", "ascii"))
seed_set = list(timgraph.get_seed_set(self.epsilon))
timgraph = None
# Simulates the chosen seed_set's performance in real world
online_spread, tried_cnts, success_cnts = self.simulate_spread(seed_set)
# Update influence estimates and counters only if at exploration epoch
if(explore_bool):
total_cost = self.active_update(tried_cnts, success_cnts, context_idx)
else:
total_cost = 0
# Oracle run
real_infs = self.context_influences(self.context_vector)
self.dump_graph(real_infs, ("tim_"+self.graph_file))
oracle = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"), self.node_cnt, self.edge_cnt, self.seed_size, bytes("IC", "ascii"))
oracle_set = list(oracle.get_seed_set(self.epsilon))
oracle = None
oracle_spread, _, _ = self.simulate_spread(oracle_set)
self.regret.append((oracle_spread + total_cost) - online_spread)
self.spread.append(online_spread)
self.update_l2_error(real_infs, inf_ests)
def run(self):
"""------------------------------------------------------------
High level function that runs the online influence maximization
algorithm for self.epochs times and reports aggregated regret
------------------------------------------------------------"""
for epoch_idx in np.arange(1, self.epochs+1):
print(epoch_idx)
self.get_context()
context_idx = self.context_classifier(self.context_vector)
under_explored = self.under_explored_nodes(context_idx, epoch_idx)
# If there are enough under-explored edges, return them
if(len(under_explored) == self.seed_size):
exploration_phase = True
seed_set = under_explored
# Otherwise, run TIM
else:
exploration_phase = False
self.dump_graph(self.inf_ests[context_idx], ("tim_"+self.graph_file))
timgraph = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"), self.node_cnt, self.edge_cnt,
(self.seed_size - len(under_explored)), bytes("IC", "ascii"))
tim_set = timgraph.get_seed_set(self.epsilon)
seed_set = list(tim_set)
seed_set.extend(under_explored)
timgraph = None
# Simulates the chosen seed_set's performance in real world
online_spread, tried_cnts, success_cnts = self.simulate_spread(seed_set)
if(exploration_phase):
total_cost = self.active_update(tried_cnts, success_cnts, context_idx, epoch_idx)
else:
total_cost = 0
# Oracle run
real_infs = self.context_influences(self.context_vector)
self.dump_graph(real_infs, ("tim_"+self.graph_file))
oracle = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"), self.node_cnt, self.edge_cnt, self.seed_size, bytes("IC", "ascii"))
oracle_set = list(oracle.get_seed_set(self.epsilon))
oracle = None
oracle_spread, _, _ = self.simulate_spread(oracle_set)
self.regret.append((oracle_spread + total_cost) - online_spread)
self.spread.append(online_spread)
self.update_l2_error(real_infs, self.inf_ests[context_idx])
def run(self):
"""------------------------------------------------------------
High level function that runs the online influence maximization
algorithm for self.epochs times and reports aggregated regret
------------------------------------------------------------"""
for epoch_idx in np.arange(1, self.epochs+1):
# Retrieve the parameters and compute inf. estimates
self.get_context()
alphas = self.local_alphas + self.global_alpha
betas = self.local_betas + self.global_beta
mus = alphas / (alphas + betas)
sigmas = (1 / (alphas + betas)) * np.sqrt((alphas * betas)/(alphas+betas+1))
inf_ests = [max(mus[idx] + (self.theta * sigmas[idx]), 0)
if(self.local_alphas[idx] + self.local_betas[idx] > 0)
else (0) for idx, _ in enumerate(self.edges[:,0])]
# Run TIM
self.dump_graph(inf_ests, ("tim_"+self.graph_file))
timgraph = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"), self.node_cnt, self.edge_cnt,
self.seed_size, bytes("IC", "ascii"))
tim_set = timgraph.get_seed_set(self.epsilon)
seed_set = list(tim_set)
timgraph = None
# Simulates the chosen seed_set's performance in real world
influenced_nodes = self.simulate_spread(seed_set)
online_spread = len(influenced_nodes)
# Update influence estimates, counters and globals
success_cnts, tried_cnts = self.random_update(influenced_nodes, seed_set)
fail_cnts = tried_cnts - success_cnts
self.update_globals(success_cnts, fail_cnts)
self.spread.append(online_spread)
self.exponentiated_gradient()
# Oracle run
real_infs = self.context_influences(self.context_vector)
self.dump_graph(real_infs, ("tim_"+self.graph_file))
oracle = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"), self.node_cnt, self.edge_cnt, self.seed_size, bytes("IC", "ascii"))
oracle_set = list(oracle.get_seed_set(self.epsilon))
oracle = None
influenced_nodes = self.simulate_spread(oracle_set)
oracle_spread = len(influenced_nodes)
self.regret.append(oracle_spread - online_spread)
self.update_l2_error(real_infs, inf_ests)
def run(self):
"""------------------------------------------------------------
High level function that runs the online influence maximization
algorithm for self.rounds times and reports aggregated regret
------------------------------------------------------------"""
for epoch_idx in np.arange(1, self.epochs+1):
if((epoch_idx+1) % 100 == 0):
print(epoch_idx+1)
self.get_context()
context_idx = self.context_classifier(self.context_vector)
under_explored = self.under_explored_nodes(context_idx, epoch_idx)
# If there are enough under-explored edges, return them
if(len(under_explored) == self.seed_size):
seed_set = under_explored
# Otherwise, run TIM
else:
self.dump_graph(self.inf_ests[context_idx], ("tim_"+self.graph_file))
timgraph = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"), self.node_cnt, self.edge_cnt,
(self.seed_size - len(under_explored)), bytes("IC", "ascii"))
tim_set = timgraph.get_seed_set(self.epsilon)
seed_set = list(tim_set)
seed_set.extend(under_explored)
timgraph = None
# Simulates the chosen seed_set's performance in real world
online_spread, tried_cnts, success_cnts = self.simulate_spread(seed_set)
# Update influence estimates and counters
self.counters[context_idx] += tried_cnts
self.successes[context_idx] += success_cnts
for edge_idx, cnt in enumerate(self.counters[context_idx]):
self.inf_ests[context_idx][edge_idx] = self.successes[context_idx][edge_idx] / cnt if(cnt > 0) else (0)
# Oracle run
real_infs = self.context_influences(self.context_vector)
self.dump_graph(real_infs, ("tim_"+self.graph_file))
oracle = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"), self.node_cnt, self.edge_cnt, self.seed_size, bytes("IC", "ascii"))
oracle_set = list(oracle.get_seed_set(self.epsilon))
oracle = None
oracle_spread, _, _ = self.simulate_spread(oracle_set)
self.regret.append(oracle_spread - online_spread)
self.spread.append(online_spread)
self.update_l2_error(real_infs, self.inf_ests[context_idx])
def run(self):
"""------------------------------------------------------------
High level function that runs the online influence maximization
algorithm for self.rounds times and reports aggregated regret
------------------------------------------------------------"""
for epoch_idx in np.arange(1, self.epochs+1):
self.get_context()
context_idx = self.context_classifier(self.context_vector)
self.dump_graph(self.inf_ests[context_idx], ("tim_"+self.graph_file))
timgraph = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"), self.node_cnt, self.edge_cnt,
(self.seed_size), bytes("IC", "ascii"))
tim_set = timgraph.get_seed_set(self.epsilon)
seed_set = list(tim_set)
timgraph = None
# Simulates the chosen seed_set's performance in real world
influenced_nodes = self.simulate_spread(seed_set)
online_spread = len(influenced_nodes)
# Update influence estimates and counters
self.random_update(context_idx, influenced_nodes, seed_set)
# Oracle run
real_infs = self.context_influences(self.context_vector)
self.dump_graph(real_infs, ("tim_"+self.graph_file))
oracle = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"), self.node_cnt, self.edge_cnt, self.seed_size, bytes("IC", "ascii"))
oracle_set = list(oracle.get_seed_set(self.epsilon))
oracle = None
influenced_nodes = self.simulate_spread(oracle_set)
oracle_spread = len(influenced_nodes)
self.regret.append(oracle_spread - online_spread)
self.spread.append(online_spread)
self.update_l2_error(real_infs, self.inf_ests[context_idx])
def run(self):
"""------------------------------------------------------------
High level function that runs the online influence maximization
algorithm for self.rounds times and reports aggregated regret
------------------------------------------------------------"""
seed_set = np.argsort(self.outdegs)[::-1][:self.seed_size].tolist()
for epoch_idx in np.arange(1, self.epochs+1):
self.get_context()
# Simulates the chosen seed_set's performance in real world
online_spread, tried_cnts, success_cnts = self.simulate_spread(seed_set)
# Oracle run
real_infs = self.context_influences(self.context_vector)
self.dump_graph(real_infs, ("tim_"+self.graph_file))
oracle = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"), self.node_cnt, self.edge_cnt, self.seed_size, bytes("IC", "ascii"))
oracle_set = list(oracle.get_seed_set(self.epsilon))
oracle = None
oracle_spread, _, _ = self.simulate_spread(oracle_set)
self.regret.append(oracle_spread - online_spread)
self.spread.append(online_spread)
def run(self):
"""------------------------------------------------------------
High level function that runs the online influence maximization
algorithm for self.epochs times and reports aggregated regret
------------------------------------------------------------"""
for epoch_idx in np.arange(1, self.epochs+1):
if((epoch_idx+1) % 100 == 0):
print(epoch_idx+1)
self.get_context()
alphas = self.local_alphas + self.global_alpha
betas = self.local_betas + self.global_beta
inf_ests = [np.random.beta(alphas[idx], betas[idx]) for idx in range(self.edge_cnt)]
self.dump_graph(inf_ests, ("tim_"+self.graph_file))
timgraph = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"), self.node_cnt, self.edge_cnt,
self.seed_size, bytes("IC", "ascii"))
tim_set = timgraph.get_seed_set(self.epsilon)
seed_set = list(tim_set)
timgraph = None
# Simulates the chosen seed_set's performance in real world
online_spread, tried_cnts, success_cnts = self.simulate_spread(seed_set)
fail_cnts = tried_cnts - success_cnts
self.local_alphas += success_cnts
self.local_betas += fail_cnts
self.update_globals(tried_cnts, success_cnts, fail_cnts)
self.spread.append(online_spread)
# Oracle run
real_infs = self.context_influences(self.context_vector)
self.dump_graph(real_infs, ("tim_"+self.graph_file))
oracle = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"), self.node_cnt, self.edge_cnt, self.seed_size, bytes("IC", "ascii"))
oracle_set = list(oracle.get_seed_set(self.epsilon))
oracle = None
oracle_spread, _, _ = self.simulate_spread(oracle_set)
self.regret.append(oracle_spread - online_spread)
self.update_l2_error(real_infs, inf_ests)
def run(self):
"""------------------------------------------------------------
High level function that runs the online influence maximization
algorithm for self.epochs times and reports aggregated regret
------------------------------------------------------------"""
for epoch_idx in np.arange(1, self.epochs + 1):
print(epoch_idx)
self.get_context()
context_idx = self.context_classifier(self.context_vector)
explore_bool = binomial(1, self.epsilons[epoch_idx - 1])
# Exploration Epoch, set influence estimates accordingly
if (explore_bool):
inf_ests = [(alpha / (alpha + beta)) * (alpha + math.sqrt(
(alpha * beta) / (alpha + beta + 1))) if
(alpha + beta > 0) else (0)
for alpha, beta in zip(self.alphas, self.betas)]
# Exploitation Epoch, set influence estimates accordingly
else:
inf_ests = [(alpha / (alpha + beta)) if (alpha + beta > 0) else
(0)
for alpha, beta in zip(self.alphas, self.betas)]
self.dump_graph(inf_ests, ("tim_" + self.graph_file))
timgraph = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"),
self.node_cnt, self.edge_cnt, self.seed_size,
bytes("IC", "ascii"))
seed_set = list(timgraph.get_seed_set(self.epsilon))
timgraph = None
# Simulates the chosen seed_set's performance in real world
online_spread, tried_cnts, success_cnts = self.simulate_spread(
seed_set)
# Update influence estimates and counters only if at exploration epoch
if (explore_bool):
total_cost = self.active_update(tried_cnts, success_cnts,
context_idx)
else:
total_cost = 0
# Oracle run
real_infs = self.context_influences(self.context_vector)
self.dump_graph(real_infs, ("tim_" + self.graph_file))
oracle = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"),
self.node_cnt, self.edge_cnt, self.seed_size,
bytes("IC", "ascii"))
oracle_set = list(oracle.get_seed_set(self.epsilon))
oracle = None
oracle_spread, _, _ = self.simulate_spread(oracle_set)
self.regret.append((oracle_spread + total_cost) - online_spread)
self.spread.append(online_spread)
self.update_l2_error(real_infs, inf_ests)
def run(self):
"""------------------------------------------------------------
High level function that runs the online influence maximization
algorithm for self.rounds times and reports aggregated regret
------------------------------------------------------------"""
for epoch_idx in np.arange(1, self.epochs + 1):
if ((epoch_idx + 1) % 100 == 0):
print(epoch_idx + 1)
self.get_context()
context_idx = self.context_classifier(self.context_vector)
under_explored = self.under_explored_nodes(context_idx, epoch_idx)
# If there are enough under-explored edges, return them
if (len(under_explored) == self.seed_size):
seed_set = under_explored
# Otherwise, run TIM
else:
self.dump_graph(self.inf_ests[context_idx],
("tim_" + self.graph_file))
timgraph = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"),
self.node_cnt, self.edge_cnt,
(self.seed_size - len(under_explored)),
bytes("IC", "ascii"))
tim_set = timgraph.get_seed_set(self.epsilon)
seed_set = list(tim_set)
seed_set.extend(under_explored)
timgraph = None
# Simulates the chosen seed_set's performance in real world
online_spread, tried_cnts, success_cnts = self.simulate_spread(
seed_set)
# Update influence estimates and counters
self.counters[context_idx] += tried_cnts
self.successes[context_idx] += success_cnts
for edge_idx, cnt in enumerate(self.counters[context_idx]):
self.inf_ests[context_idx][edge_idx] = self.successes[
context_idx][edge_idx] / cnt if (cnt > 0) else (0)
# Oracle run
real_infs = self.context_influences(self.context_vector)
self.dump_graph(real_infs, ("tim_" + self.graph_file))
oracle = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"),
self.node_cnt, self.edge_cnt, self.seed_size,
bytes("IC", "ascii"))
oracle_set = list(oracle.get_seed_set(self.epsilon))
oracle = None
oracle_spread, _, _ = self.simulate_spread(oracle_set)
self.regret.append(oracle_spread - online_spread)
self.spread.append(online_spread)
self.update_l2_error(real_infs, self.inf_ests[context_idx])
def run(self):
"""------------------------------------------------------------
High level function that runs the online influence maximization
algorithm for self.epochs times and reports aggregated regret
------------------------------------------------------------"""
for epoch_idx in np.arange(1, self.epochs + 1):
print(epoch_idx)
self.get_context()
context_idx = self.context_classifier(self.context_vector)
under_explored = self.under_explored_nodes(context_idx, epoch_idx)
# If there are enough under-explored edges, return them
if (len(under_explored) == self.seed_size):
exploration_phase = True
seed_set = under_explored
# Otherwise, run TIM
else:
exploration_phase = False
self.dump_graph(self.inf_ests[context_idx],
("tim_" + self.graph_file))
timgraph = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"),
self.node_cnt, self.edge_cnt,
(self.seed_size - len(under_explored)),
bytes("IC", "ascii"))
tim_set = timgraph.get_seed_set(self.epsilon)
seed_set = list(tim_set)
seed_set.extend(under_explored)
timgraph = None
# Simulates the chosen seed_set's performance in real world
online_spread, tried_cnts, success_cnts = self.simulate_spread(
seed_set)
if (exploration_phase):
total_cost = self.active_update(tried_cnts, success_cnts,
context_idx, epoch_idx)
else:
total_cost = 0
# Oracle run
real_infs = self.context_influences(self.context_vector)
self.dump_graph(real_infs, ("tim_" + self.graph_file))
oracle = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"),
self.node_cnt, self.edge_cnt, self.seed_size,
bytes("IC", "ascii"))
oracle_set = list(oracle.get_seed_set(self.epsilon))
oracle = None
oracle_spread, _, _ = self.simulate_spread(oracle_set)
self.regret.append((oracle_spread + total_cost) - online_spread)
self.spread.append(online_spread)
self.update_l2_error(real_infs, self.inf_ests[context_idx])
def run(self):
"""------------------------------------------------------------
High level function that runs the online influence maximization
algorithm for self.epochs times and reports aggregated regret
------------------------------------------------------------"""
for epoch_idx in np.arange(1, self.epochs + 1):
# Retrieve the parameters and compute inf. estimates
self.get_context()
alphas = self.local_alphas + self.global_alpha
betas = self.local_betas + self.global_beta
mus = alphas / (alphas + betas)
sigmas = (1 / (alphas + betas)) * np.sqrt(
(alphas * betas) / (alphas + betas + 1))
inf_ests = [
max(mus[idx] + (self.theta * sigmas[idx]), 0) if
(self.local_alphas[idx] + self.local_betas[idx] > 0) else (0)
for idx, _ in enumerate(self.edges[:, 0])
]
# Run TIM
self.dump_graph(inf_ests, ("tim_" + self.graph_file))
timgraph = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"),
self.node_cnt, self.edge_cnt, self.seed_size,
bytes("IC", "ascii"))
tim_set = timgraph.get_seed_set(self.epsilon)
seed_set = list(tim_set)
timgraph = None
# Simulates the chosen seed_set's performance in real world
online_spread, tried_cnts, success_cnts = self.simulate_spread(
seed_set)
fail_cnts = tried_cnts - success_cnts
self.local_alphas += success_cnts
self.local_betas += fail_cnts
self.update_globals(tried_cnts, success_cnts, fail_cnts)
self.spread.append(online_spread)
self.exponentiated_gradient()
# Oracle run
real_infs = self.context_influences(self.context_vector)
self.dump_graph(real_infs, ("tim_" + self.graph_file))
oracle = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"),
self.node_cnt, self.edge_cnt, self.seed_size,
bytes("IC", "ascii"))
oracle_set = list(oracle.get_seed_set(self.epsilon))
oracle = None
oracle_spread, _, _ = self.simulate_spread(oracle_set)
self.regret.append(oracle_spread - online_spread)
self.update_l2_error(real_infs, inf_ests)
def run(self):
"""------------------------------------------------------------
High level function that runs the online influence maximization
algorithm for self.epochs times and reports aggregated regret
------------------------------------------------------------"""
for epoch_idx in np.arange(1, self.epochs + 1):
if ((epoch_idx + 1) % 100 == 0):
print(epoch_idx + 1)
self.get_context()
alphas = self.local_alphas + self.global_alpha
betas = self.local_betas + self.global_beta
inf_ests = [
np.random.beta(alphas[idx], betas[idx])
for idx in range(self.edge_cnt)
]
self.dump_graph(inf_ests, ("tim_" + self.graph_file))
timgraph = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"),
self.node_cnt, self.edge_cnt, self.seed_size,
bytes("IC", "ascii"))
tim_set = timgraph.get_seed_set(self.epsilon)
seed_set = list(tim_set)
timgraph = None
# Simulates the chosen seed_set's performance in real world
online_spread, tried_cnts, success_cnts = self.simulate_spread(
seed_set)
fail_cnts = tried_cnts - success_cnts
self.local_alphas += success_cnts
self.local_betas += fail_cnts
self.update_globals(tried_cnts, success_cnts, fail_cnts)
self.spread.append(online_spread)
# Oracle run
real_infs = self.context_influences(self.context_vector)
self.dump_graph(real_infs, ("tim_" + self.graph_file))
oracle = PyTimGraph(bytes("tim_" + self.graph_file, "ascii"),
self.node_cnt, self.edge_cnt, self.seed_size,
bytes("IC", "ascii"))
oracle_set = list(oracle.get_seed_set(self.epsilon))
oracle = None
oracle_spread, _, _ = self.simulate_spread(oracle_set)
self.regret.append(oracle_spread - online_spread)
self.update_l2_error(real_infs, inf_ests)