def plot(ground_truth, estimation_exp, estimator_exp, estimation_mnl, estimator_mnl, transactions):
print('')
print('Log-likelihood RND: %.4f' % RandomChoiceModel.simple_random(ground_truth.products).log_likelihood_for(transactions))
print('Log-likelihood EXP: %.4f' % estimation_exp.log_likelihood_for(transactions))
print('Log-likelihood MNL: %.4f' % estimation_mnl.log_likelihood_for(transactions))
print('')
print('Soft RMSE RND: %.4f' % ground_truth.soft_rmse_for(RandomChoiceModel.simple_random(ground_truth.products)))
print('Soft RMSE EXP: %.4f' % ground_truth.soft_rmse_for(estimation_exp))
print('Soft RMSE MNL: %.4f' % ground_truth.soft_rmse_for(estimation_mnl))
print('')
iterations_exp = estimator_exp.profiler().iterations()
iterations_mnl = estimator_mnl.profiler().iterations()
y_data_exp = [-i.value() for i in iterations_exp]
x_data_exp = [i.stop_time() - iterations_exp[0].start_time() for i in iterations_exp]
y_data_mnl = [-i.value() for i in iterations_mnl]
x_data_mnl = [i.stop_time() - iterations_mnl[0].start_time() for i in iterations_mnl]
plt.plot(x_data_exp, y_data_exp, x_data_mnl, y_data_mnl)
plt.title('MNL vs EXP (Ground truth MNL)')
plt.xlabel('Time (seconds)')
plt.ylabel('Log-likelihood')
plt.legend(['exp', 'mnl'])
plt.savefig('output/MNL vs EXP (Ground truth MNL).png')
def plot(ground_truth, estimator_max, estimator_em, estimator_mkv2,
estimation_max, estimation_em, estimation_mkv2, transactions):
print('')
print('Log-likelihood RND: %.4f' % RandomChoiceModel.simple_random(
ground_truth.products).log_likelihood_for(transactions))
print('Log-likelihood MAX: %.4f' %
estimation_max.log_likelihood_for(transactions))
print('Log-likelihood EM: %.4f' %
estimation_em.log_likelihood_for(transactions))
print('Log-likelihood R2: %.4f' %
estimation_mkv2.log_likelihood_for(transactions))
print('')
print('Soft RMSE RND: %.4f' % ground_truth.soft_rmse_for(
RandomChoiceModel.simple_random(ground_truth.products)))
print('Soft RMSE MAX: %.4f' % ground_truth.soft_rmse_for(estimation_max))
print('Soft RMSE EM: %.4f' % ground_truth.soft_rmse_for(estimation_em))
print('Soft RMSE R2: %.4f' % ground_truth.soft_rmse_for(estimation_mkv2))
print('')
iterations_max = estimator_max.profiler().iterations()[2:]
iterations_em = estimator_em.profiler().iterations()[2:]
iterations_mkv2 = estimator_mkv2.profiler().iterations()[2:]
y_mkv2_data = [-i.value() for i in iterations_mkv2]
x_mkv2_data = [
i.stop_time() - iterations_mkv2[0].start_time()
for i in iterations_mkv2
]
y_max_data = [-i.value() for i in iterations_max]
x_max_data = [
i.stop_time() - iterations_max[0].start_time() for i in iterations_max
]
y_em_data = [i.value() for i in iterations_em]
x_em_data = [
i.stop_time() - iterations_em[0].start_time() for i in iterations_em
]
plt.plot(x_max_data, y_max_data, x_em_data, y_em_data, x_mkv2_data,
y_mkv2_data)
plt.title('Markov Chain method comparison')
plt.xlabel('Time (seconds)')
plt.ylabel('Log-likelihood')
plt.legend(['max', 'em', 'rank2'])
plt.savefig('output/Markov Chain method comparison.png')
def main():
Settings.new(
linear_solver_partial_time_limit=300.0,
non_linear_solver_partial_time_limit=300.0,
solver_total_time_limit=1800.0,
)
ground_truth, transactions = read_synthetic_instance(
'instances/100_lists_300_periods_1_instance.json')
products = ground_truth.products
nests = [{
'products': [0],
'lambda': 0.8
}, {
'products': [i for i in range(1, len(products)) if i % 2 == 0],
'lambda': 0.8
}, {
'products': [i for i in range(1, len(products)) if i % 2 == 1],
'lambda': 0.8
}]
lists = [[i] + list(sorted(set(products) - {i}))
for i in range(len(products))]
models_to_run = [
(ExponomialModel.simple_deterministic(products),
MaximumLikelihoodEstimator()),
(MultinomialLogitModel.simple_deterministic(products),
MaximumLikelihoodEstimator()),
(RandomChoiceModel.simple_deterministic(products),
RandomChoiceModelMaximumLikelihoodEstimator()),
(LatentClassModel.simple_deterministic(products, 1),
LatentClassFrankWolfeEstimator()),
(MarkovChainModel.simple_deterministic(products),
MarkovChainExpectationMaximizationEstimator()),
(MarkovChainRank2Model.simple_deterministic(products),
MaximumLikelihoodEstimator()),
(MixedLogitModel.simple_deterministic(products),
MaximumLikelihoodEstimator()),
(NestedLogitModel.simple_deterministic(products, nests=nests),
MaximumLikelihoodEstimator()),
(RankedListModel.simple_deterministic(products, ranked_lists=lists),
RankedListExpectationMaximizationEstimator(MIPMarketExplorer())),
]
results = []
for initial_solution, estimator in models_to_run:
if hasattr(estimator, 'estimate_with_market_discovery'):
model = estimator.estimate_with_market_discovery(
initial_solution, transactions)
else:
model = estimator.estimate(initial_solution, transactions)
soft_rmse = model.soft_rmse_for(ground_truth)
results.append((model, estimator.profiler(), soft_rmse))
plot(results)
def plot(ground_truth, estimator_max, estimator_em, estimator_mip,
estimation_max, estimation_em, estimation_mip, transactions):
print('')
print('Log-likelihood RND: %.4f' % RandomChoiceModel.simple_random(
ground_truth.products).log_likelihood_for(transactions))
print('Log-likelihood MAX: %.4f' %
estimation_max.log_likelihood_for(transactions))
print('Log-likelihood EM: %.4f' %
estimation_em.log_likelihood_for(transactions))
print('Log-likelihood MIP: %.4f' %
estimation_mip.log_likelihood_for(transactions))
print('')
print('Soft RMSE RND: %.4f' % ground_truth.soft_rmse_for(
RandomChoiceModel.simple_random(ground_truth.products)))
print('Soft RMSE MAX: %.4f' % ground_truth.soft_rmse_for(estimation_max))
print('Soft RMSE EM: %.4f' % ground_truth.soft_rmse_for(estimation_em))
print('Soft RMSE MIP: %.4f' % ground_truth.soft_rmse_for(estimation_mip))
print('')
iterations_em = estimator_em.profiler().iterations()
iterations_max = estimator_max.profiler().iterations()
iterations_mip = estimator_mip.profiler().iterations()
y_data_max = [-i.value() for i in iterations_max]
x_data_max = [
i.stop_time() - iterations_max[0].start_time() for i in iterations_max
]
y_data_em = [i.value() for i in iterations_em]
x_data_em = [
i.stop_time() - iterations_em[0].start_time() for i in iterations_em
]
y_data_mip = [i.value() for i in iterations_mip]
x_data_mip = [
i.stop_time() - iterations_mip[0].start_time() for i in iterations_mip
]
plt.plot(x_data_max, y_data_max, x_data_em, y_data_em, x_data_mip,
y_data_mip)
plt.title('Ranked List Method comparison')
plt.xlabel('Time (seconds)')
plt.ylabel('Log-likelihood')
plt.legend(['max (has lists)', 'em (has lists)', 'mip (discovers lists)'])
plt.savefig('output/Ranked List Method comparison.png')
def plot(ground_truth, estimation_em, estimation_max, estimation_fw,
estimator_em, estimator_max, estimator_fw, transactions):
print('')
print('Log-likelihood RND: %.4f' % RandomChoiceModel.simple_random(
ground_truth.products).log_likelihood_for(transactions))
print('Log-likelihood MAX: %.4f' %
estimation_max.log_likelihood_for(transactions))
print('Log-likelihood EM: %.4f' %
estimation_em.log_likelihood_for(transactions))
print('Log-likelihood FW: %.4f' %
estimation_fw.log_likelihood_for(transactions))
print('')
print('Soft RMSE RND: %.4f' % ground_truth.soft_rmse_for(
RandomChoiceModel.simple_random(ground_truth.products)))
print('Soft RMSE MAX: %.4f' % ground_truth.soft_rmse_for(estimation_max))
print('Soft RMSE EM: %.4f' % ground_truth.soft_rmse_for(estimation_em))
print('Soft RMSE FW: %.4f' % ground_truth.soft_rmse_for(estimation_fw))
print('')
iterations_fw = estimator_fw.profiler().iterations()
iterations_max = estimator_max.profiler().iterations()
iterations_em = estimator_em.profiler().iterations()
# Frank wolfe has two subproblems on each iteration.
# - 1st problem finds new MNLs to add to the Latent class: (always negative objective function)
# - 2nd problem corrects weights for each class (always positive objective function, basically a NLL)
# Not very elegant:
# - Plot only NLL after each iteration.
nlls_iterations = []
for i in range(1, len(iterations_fw)):
# If went from positive to negative, then changed from 2nd problem to 1st problem. Save NLL.
if iterations_fw[i - 1].value() >= 0 and iterations_fw[i].value() < 0:
nlls_iterations.append((-iterations_fw[i - 1].value(),
iterations_fw[i - 1].stop_time() -
iterations_fw[0].start_time()))
if iterations_fw[-1].value() >= 0:
nlls_iterations.append(
(-iterations_fw[-1].value(),
iterations_fw[-1].stop_time() - iterations_fw[0].start_time()))
y_data_fw = [
x[0] * len(set(transactions)) / len(transactions)
for x in nlls_iterations
]
x_data_fw = [x[1] for x in nlls_iterations]
y_data_max = [-i.value() for i in iterations_max]
x_data_max = [
i.stop_time() - iterations_max[0].start_time() for i in iterations_max
]
y_data_em = [i.value() for i in iterations_em]
x_data_em = [
i.stop_time() - iterations_em[0].start_time() for i in iterations_em
]
plt.plot(x_data_fw, y_data_fw, x_data_max, y_data_max, x_data_em,
y_data_em)
plt.title('Latent Class method comparison')
plt.xlabel('Time (seconds)')
plt.ylabel('Log-likelihood')
plt.legend([
'CG (%s classes discovered)' % len(estimation_fw.gammas),
'MAX (%s classes given)' % len(estimation_max.gammas),
'EM (%s classes given)' % len(estimation_em.gammas)
])
plt.savefig('output/Latent Class method comparison.png')
'mx': {
'estimator':
MaximumLikelihoodEstimator(),
'model_class':
lambda products: MixedLogitModel.simple_deterministic(products
),
'name':
'Mixed Logit',
'settings':
NORMAL_SETTINGS
},
'rnd': {
'estimator':
RandomChoiceModelMaximumLikelihoodEstimator(),
'model_class':
lambda products: RandomChoiceModel.simple_deterministic(
products),
'name':
'Random Choice',
'settings':
NORMAL_SETTINGS
},
'rl': {
'estimator':
RankedListMaximumLikelihoodEstimator.with_this(
MIPMarketExplorer()),
'model_class':
lambda products: RankedListModel.
simple_deterministic_independent(products),
'name':
'Ranked List',
'settings':