def event_type_map_eval_ml_metrics(config, logger, result):
print_and_log(logger, "====================================")
print_and_log(logger, "event type MAP evaluation:")
print_and_log(
logger, "event type MAP: {}".format(
ml_metrics.mapk(result['gt_all_event_id'],
result['pred_all_event_id'])))
k_list = []
for i in range(len(result['chain_name'])):
k_list.append(len(result['gt_all_event_type'][i]))
k_list = sorted(list(set(k_list)))
k_list.remove(0)
print_and_log(logger, "all possible k: {}".format(k_list))
for k in k_list:
map_at_k = ml_metrics.mapk(result['gt_all_event_id'],
result['pred_all_event_id'], k)
print_and_log(logger, "event type MAP@{}: {}".format(int(k), map_at_k))
return result
def event_type_map_eval_given_gt(config, logger, result):
print_and_log(logger, "====================================")
print_and_log(logger, "event type MAP evaluation:")
result['et_ap'] = list()
for a, p in zip(result['gt_all_event_id'], result['pred_all_event_id']):
AP = precision_score(a, p, average='macro')
result['et_ap'].append(AP)
map_re = np.mean(result['et_ap'])
result['et_map'] = map_re
print_and_log(logger, "event type MAP: {}".format(round(map_re, 4)))
return result
def event_type_map_eval(config, logger, result):
print_and_log(logger, "====================================")
print_and_log(logger, "event type MAP evaluation:")
result['et_ap'] = list()
for a, p in zip(result['gt_all_event_id'], result['pred_all_event_id']):
AP = compute_AP(a, p)
result['et_ap'].append(AP)
map_re = np.mean(result['et_ap'])
result['et_map'] = map_re
print_and_log(logger, "event type MAP: {}".format(round(map_re, 4)))
return result
def user_cluster_map_eval(config, logger, result):
print_and_log(logger, "====================================")
print_and_log(logger, "user cluster MAP evaluation:")
result['uc_ap'] = list()
for a, p in zip(result['gt_all_user_cluster'],
result['pred_all_user_cluster']):
AP = compute_AP(a, p)
result['uc_ap'].append(AP)
map_re = np.mean(result['uc_ap'])
result['uc_map'] = map_re
print_and_log(logger, "user cluster MAP: {}".format(round(map_re, 4)))
return result
def event_type_categorical_accuracy_eval_given_gt(config, logger, result):
print_and_log(logger, "====================================")
print_and_log(logger, "event type categorical accuracy evaluation:")
y_true = []
y_pred = []
for a, p in zip(result['gt_all_event_id'], result['pred_all_event_id']):
y_true += a
y_pred += p
result['et_cate'] = accuracy_score(y_true, y_pred)
print_and_log(
logger, "event type categorical accuracy: {}".format(
round(result['et_cate'], 4)))
return result
def user_cluster_categorical_accuracy_eval_given_gt(config, logger, result):
print_and_log(logger, "====================================")
print_and_log(logger, "user cluster categorical accuracy evaluation:")
y_true = []
y_pred = []
for a, p in zip(result['gt_all_user_cluster'],
result['pred_all_user_cluster']):
y_true += a
y_pred += p
result['uc_cate'] = accuracy_score(y_true, y_pred)
print_and_log(
logger, "user cluster categorical accuracy: {}".format(
round(result['uc_cate'], 4)))
return result
def time_delay_overall_evaluation(config,
logger,
result,
result_save_path,
plot_ts=True,
chain_length_eval=True):
print_and_log(logger, "====================================")
print_and_log(logger, "time delay evaluation:")
# statistics
pred_all = []
gt_all = []
avg_dtw = []
avg_mse = []
result["td_DTW"] = list()
result["td_MSE"] = list()
for i in range(len(result['chain_name'])):
pred_time_delay = result['pred_all_time_delay'][i]
gt_time_delay = result['gt_all_time_delay'][i]
if len(pred_time_delay) == 0:
pred_time_delay = [-1]
if len(gt_time_delay) == 0:
gt_time_delay = [-1]
avg_dtw.append(fastdtw(gt_time_delay, pred_time_delay)[0])
result["td_DTW"].append(avg_dtw[-1])
if len(gt_time_delay) == len(pred_time_delay):
avg_mse.append(mean_squared_error(gt_time_delay, pred_time_delay))
result["td_MSE"].append(avg_mse[-1])
else:
result["td_MSE"].append('null')
if len(result['pred_all_time_delay'][i]) != 0:
pred_all += pred_time_delay
if len(result['gt_all_time_delay'][i]) != 0:
gt_all += gt_time_delay
print_and_log(logger, "Average DTW: {}".format(round(np.mean(avg_dtw), 4)))
if config['given_gt']:
print_and_log(logger, "Average MSE: {}".format(np.mean(avg_mse)))
print_and_log(
logger,
"MAX predicted: {}, ground truth: {}".format(round(max(pred_all), 4),
round(max(gt_all), 4)))
print_and_log(
logger,
"MIN predicted: {}, ground truth: {}".format(round(min(pred_all), 4),
round(min(gt_all), 4)))
print_and_log(
logger, "MEAN predicted: {}, ground truth: {}".format(
round(np.mean(pred_all), 4), round(np.mean(gt_all), 4)))
print_and_log(
logger, "STD predicted: {}, ground truth: {}".format(
round(np.std(pred_all), 4), round(np.std(gt_all), 4)))
# chain length evaluation
if chain_length_eval:
length_mae = []
length_stat = dict()
length_stat["gt_chain_0"] = 0
length_stat["gt_chain_1"] = 0
length_stat["Same_as_gt"] = 0
length_stat["diff_1_to_10"] = 0
length_stat["diff_10_to_100"] = 0
length_stat["diff_100+"] = 0
if 'chains_applied_keep_pred' in result:
length_stat["applied_threshold"] = len(
result["chains_applied_keep_pred"])
sim_start = config['sim_period']['start'].split('T')[0]
sim_end = config['sim_period']['end'].split('T')[0]
if plot_ts:
time_delay_plot_save_path = os.path.join(result_save_path,
"time_delay_plot")
if not os.path.exists(time_delay_plot_save_path):
os.makedirs(time_delay_plot_save_path)
if chain_length_eval or plot_ts:
for i in range(len(result['chain_name'])):
chain = result['chain_name'][i]
pred_time_delay = result['pred_all_time_delay'][i]
gt_time_delay = result['gt_all_time_delay'][i]
if plot_ts:
plot_time_delay_ts_for_one_chain(chain,
time_delay_plot_save_path,
pred_time_delay,
gt_time_delay, sim_start,
sim_end)
if chain_length_eval:
length_mae.append(
abs(len(pred_time_delay) - len(gt_time_delay)))
if len(gt_time_delay) == 0:
length_stat["gt_chain_0"] += 1
if len(gt_time_delay) == 1:
length_stat["gt_chain_1"] += 1
if len(pred_time_delay) == len(gt_time_delay):
length_stat["Same_as_gt"] += 1
if abs(len(pred_time_delay) - len(gt_time_delay)) < 10 and (
abs(len(pred_time_delay) - len(gt_time_delay)) >= 1):
length_stat["diff_1_to_10"] += 1
if abs(len(pred_time_delay) - len(gt_time_delay)) < 100 and (
abs(len(pred_time_delay) - len(gt_time_delay)) >= 10):
length_stat["diff_10_to_100"] += 1
if abs(len(pred_time_delay) - len(gt_time_delay)) >= 100:
length_stat["diff_100+"] += 1
if chain_length_eval:
length_mae = np.mean(length_mae)
if chain_length_eval:
print_and_log(logger, "====================================")
print_and_log(logger, "chain length evaluation:")
print_and_log(logger, "MAE: {}".format(round(length_mae, 4)))
print_and_log(
logger, "Count of number of simulated "
"chains: {}".format(len(result['chain_name'])))
print_and_log(
logger, "Count of number of chains whose "
"ground truth length is 0: {}".format(length_stat["gt_chain_0"]))
print_and_log(
logger, "Count of number of chains whose "
"ground truth length is 1: {}".format(length_stat["gt_chain_1"]))
if 'chains_applied_keep_pred' in result:
print_and_log(
logger, "Count of number of predicted chains that "
"length needed threshold to be applied: {}, "
"percentage: {} ".format(
length_stat["applied_threshold"],
round(
length_stat["applied_threshold"] /
len(result['chain_name']), 4)))
print_and_log(
logger, "Count of number of predicted "
"chains that has "
"same length as ground truth"
": {}, percentage: {}".format(
length_stat["Same_as_gt"],
round(length_stat["Same_as_gt"] / len(result['chain_name']),
4)))
print_and_log(
logger, "Count of number of predicted chains that "
"length difference is 1 to 10: {},"
"percentage: {}".format(
length_stat["diff_1_to_10"],
round(length_stat["diff_1_to_10"] / len(result['chain_name']),
4)))
print_and_log(
logger, "Count of number of predicted chains that "
"length difference is 10 to 100: {}, "
"percentage: {}".format(
length_stat["diff_10_to_100"],
round(
length_stat["diff_10_to_100"] / len(result['chain_name']),
4)))
print_and_log(
logger, "Count of number of predicted chains that "
"length difference is 100 and above: {}, "
"percentage: {}".format(
length_stat["diff_100+"],
round(length_stat["diff_100+"] / len(result['chain_name']),
4)))
return result
def user_cluster_nlg_eval(config, logger, result):
print_and_log(logger, "====================================")
print_and_log(logger, "user cluster average bleu scores:")
# print_and_log(logger, "Please install nlg-eval package!\n"
# "Reference: https://github.com/Maluuba/nlg-eval")
# print_and_log(logger, "After installing, please change the package "
# "__init__.py file (contact: [email protected]).")
sys.path.append(config['nlgeval_repo_dir'])
from nlgeval import compute_individual_metrics
# avg bleu
avg_bleu = dict()
avg_bleu = dict()
avg_bleu['Bleu_1'] = list()
avg_bleu['Bleu_2'] = list()
avg_bleu['Bleu_3'] = list()
avg_bleu['Bleu_4'] = list()
result['uc_bleu1'] = list()
result['uc_bleu2'] = list()
result['uc_bleu3'] = list()
result['uc_bleu4'] = list()
for i in range(len(result['chain_name'])):
if len(result['gt_all_user_cluster'][i]) == 0:
gt_chain = " ".join(['no_event_in_simperiod'])
else:
gt_chain = " ".join(
[str(ele) for ele in result['gt_all_user_cluster'][i]])
if len(result['pred_all_user_cluster'][i]) == 0:
hy_chain = " ".join(['no_event_in_simperiod'])
else:
hy_chain = " ".join(
[str(ele) for ele in result['pred_all_user_cluster'][i]])
metrics_dict = compute_individual_metrics(gt_chain,
hy_chain,
no_overlap=(False, True),
no_skipthoughts=True,
no_glove=True)
result['uc_bleu1'].append(metrics_dict['Bleu_1'])
avg_bleu['Bleu_1'].append(metrics_dict['Bleu_1'])
if len(result['gt_all_user_cluster'][i]) >= 2: # and (
# len(result['pred_all_user_cluster'][i]) >= 2
# ):
result['uc_bleu2'].append(metrics_dict['Bleu_2'])
avg_bleu['Bleu_2'].append(metrics_dict['Bleu_2'])
else:
result['uc_bleu2'].append('null')
if len(result['gt_all_user_cluster'][i]) >= 3: # and (
# len(result['pred_all_user_cluster'][i])
# ):
result['uc_bleu3'].append(metrics_dict['Bleu_3'])
avg_bleu['Bleu_3'].append(metrics_dict['Bleu_3'])
else:
result['uc_bleu3'].append('null')
if len(result['gt_all_user_cluster'][i]) >= 4: # and (
# len(result['pred_all_user_cluster'][i]) >= 4
# ):
result['uc_bleu4'].append(metrics_dict['Bleu_4'])
avg_bleu['Bleu_4'].append(metrics_dict['Bleu_4'])
else:
result['uc_bleu4'].append('null')
for metric in avg_bleu:
print_and_log(
logger, "{}: {}".format(metric,
round(np.average(avg_bleu[metric]), 4)))
# print_and_log(logger, "{}: {}, calculated from {} values".format(
# metric, round(np.average(avg_bleu[metric]), 4),
# len(avg_bleu[metric])))
# pdb.set_trace()
return result
def result_evaluation_given_gt(config, result_save_path, only_has_event=False):
# load result pickle
with open(os.path.join(result_save_path, 'result.pickle'), 'rb') as handle:
result = pickle.load(handle)
print('result.pickle loaded!')
print("result.keys: {}\n\n".format(result.keys()))
if only_has_event:
result_new = dict()
result_new['chain_name'] = list()
result_new['pred_all_event_id'] = list()
result_new['pred_all_event_type'] = list()
result_new['pred_all_time_delay'] = list()
result_new['pred_all_user_cluster'] = list()
result_new['gt_all_event_id'] = list()
result_new['gt_all_event_type'] = list()
result_new['gt_all_time_delay'] = list()
result_new['gt_all_user_cluster'] = list()
for i in range(len(result['chain_name'])):
if len(result['gt_all_event_id'][i]) != 0:
result_new['chain_name'].append(result['chain_name'][i])
result_new['pred_all_event_id'].append(
result['pred_all_event_id'][i])
result_new['pred_all_event_type'].append(
result['pred_all_event_type'][i])
result_new['pred_all_time_delay'].append(
result['pred_all_time_delay'][i])
result_new['pred_all_user_cluster'].append(
result['pred_all_user_cluster'][i])
result_new['gt_all_event_id'].append(
result['gt_all_event_id'][i])
result_new['gt_all_event_type'].append(
result['gt_all_event_type'][i])
result_new['gt_all_time_delay'].append(
result['gt_all_time_delay'][i])
result_new['gt_all_user_cluster'].append(
result['gt_all_user_cluster'][i])
result = result_new
# logger
if only_has_event:
logger = set_logger(
os.path.join(
result_save_path, 'evaluate_only_has_event_given_gt_' +
dt.now().strftime("%Y-%m-%dT%H-%M-%SZ") + '.log'))
else:
logger = set_logger(
os.path.join(
result_save_path, 'evaluate_all_given_gt_' +
dt.now().strftime("%Y-%m-%dT%H-%M-%SZ") + '.log'))
print_and_log(
logger, "Evaluation over {} simulated chains...".format(
len(result['chain_name'])))
# evaluation proceses
if config['event_type_nlg_eval']:
result = event_type_nlg_eval(config, logger, result)
if config['event_type_map_eval']:
result = event_type_map_eval_given_gt(config, logger, result)
result = event_type_categorical_accuracy_eval_given_gt(
config, logger, result)
if config['event_type_percentage_eval']:
result = event_type_percentage_eval(config, logger, result)
if config['user_cluster_nlg_eval']:
result = user_cluster_nlg_eval(config, logger, result)
if config['user_cluster_map_eval']:
result = user_cluster_map_eval_given_gt(config, logger, result)
result = user_cluster_categorical_accuracy_eval_given_gt(
config, logger, result)
if config['user_cluster_percentage_eval']:
result = user_cluster_percentage_eval(config, logger, result)
if config['time_delay_overall_evaluation']:
if not only_has_event:
result = time_delay_overall_evaluation(config,
logger,
result,
result_save_path,
plot_ts=config['plot_ts'],
chain_length_eval=False)
else:
result = time_delay_overall_evaluation(config,
logger,
result,
result_save_path,
plot_ts=False,
chain_length_eval=False)
write_result_to_file(config, result, logger)
del logger
return
def user_cluster_percentage_eval(config, logger, result):
print_and_log(logger, "====================================")
print_and_log(logger, "user cluster distribution evaluation:")
gt_class_list = []
pred_class_list = []
for i in range(len(result['chain_name'])):
gt_class_list += result['gt_all_user_cluster'][i]
pred_class_list += result['pred_all_user_cluster'][i]
gt_class_list_counter = Counter(gt_class_list)
pred_class_list_counter = Counter(pred_class_list)
clusters = list(range(100 + 1))
counts_per_class = []
for i in range(len(clusters)):
counts_per_class.append(gt_class_list_counter[i])
gt_distribution = cal_distribution(counts_per_class)
counts_per_class = []
for i in range(len(clusters)):
counts_per_class.append(pred_class_list_counter[i])
pred_distribution = cal_distribution(counts_per_class)
print_and_log(logger, "!!!! ground truth distribution: ")
for i in range(len(clusters)):
print_and_log(logger, "{}: {}".format(i, round(gt_distribution[i], 4)))
print_and_log(logger, "!!!! prediction distribution: ")
for i in range(len(clusters)):
print_and_log(logger, "{}: {}".format(i, round(pred_distribution[i],
4)))
return result
def event_type_percentage_eval(config, logger, result):
print_and_log(logger, "====================================")
print_and_log(logger, "event type distribution evaluation:")
gt_class_list = []
pred_class_list = []
for i in range(len(result['chain_name'])):
gt_class_list += result['gt_all_event_type'][i]
pred_class_list += result['pred_all_event_type'][i]
gt_class_list_counter = Counter(gt_class_list)
pred_class_list_counter = Counter(pred_class_list)
eventtype_2_id = dict()
for key in config['eventtype_2_id']:
eventtype_2_id[key] = config['eventtype_2_id'][key] - 1
id_2_eventtype = dict(zip(eventtype_2_id.values(), eventtype_2_id.keys()))
# pdb.set_trace()
counts_per_class = []
for i in range(len(id_2_eventtype)):
et = id_2_eventtype[i]
counts_per_class.append(gt_class_list_counter[et])
gt_distribution = cal_distribution(counts_per_class)
counts_per_class = []
for i in range(len(id_2_eventtype)):
et = id_2_eventtype[i]
counts_per_class.append(pred_class_list_counter[et])
pred_distribution = cal_distribution(counts_per_class)
print_and_log(logger, "!!!! ground truth distribution: ")
for i in range(len(id_2_eventtype)):
et = id_2_eventtype[i]
print_and_log(logger, "{}: {}".format(et, round(gt_distribution[i],
4)))
print_and_log(logger, "!!!! prediction distribution: ")
for i in range(len(id_2_eventtype)):
et = id_2_eventtype[i]
print_and_log(logger, "{}: {}".format(et,
round(pred_distribution[i], 4)))
return result
def testset_creation_given_gt(config, GT_avail=True): # not implemented!
create_dataset_start = time.time()
logger = set_logger(
os.path.join(
config['exp_save_dir'], 'testset_creation_given_gt_' +
dt.now().strftime("%Y-%m-%dT%H-%M-%SZ") + '.log'))
print_and_log(logger, '{}'.format('testset creation given gt...'))
testset = dict()
testset['X_test'] = list()
testset['Y_test'] = list()
testset['repo_list'] = list()
testset['repo2sampleid'] = list()
eventtype_2_id = config['eventtype_2_id']
# id_2_eventtype = dict(zip(eventtype_2_id.values(),
# eventtype_2_id.keys()))
train_start = utc_timestamp(config['train_period']['start'])
sim_start = utc_timestamp(config['sim_period']['start'])
sim_end = utc_timestamp(config['sim_period']['end'])
with open(
os.path.join(config['exp_save_dir'], "dataset",
'unique_repo_train_vali.json'), 'r') as f:
unique_repo_train_vali = json.load(f)
repo_list = list(unique_repo_train_vali.keys())
testset['repo_list'] = repo_list
with open(config['github_event_repos_path'], 'rb') as f:
github_event_repos = pickle.load(f)
github_event_repos_set = set(
[repo[:22] + '-' + repo[23:] for repo in github_event_repos])
x_dim = config['x_dim']
if config['use_repo_embedding']:
with open(config['embed_vec_path'], 'rb') as f:
embed_vec_all = pickle.load(f)
if config['use_repo_profile_features']:
with open(config['repo_profile_feat_path'], 'r') as f:
repo_profile_raw = json.load(f)
repo_profile_feat = dict()
for repo in repo_profile_raw:
this_repo_profile_feat = []
this_repo_profile_feat.append(
time_delay_normalization(repo_profile_raw[repo]['language'][0],
'10log10_xplus1'))
this_repo_profile_feat += repo_profile_raw[repo]['user_type']
repo_profile_feat[repo] = this_repo_profile_feat
if config['use_repo_idx_features']:
with open(config['load_repo_path'], 'rb') as f:
load_repo_train_vali = pickle.load(f)
if config['use_repo_activity_features']:
max_sum_act_feat = 0
max_sum_act_feat_vec = None
with open(config['repo_activity_feat_path'], 'rb') as f:
repo_act_feat = pickle.load(f)
for repo in repo_act_feat:
for cluster in repo_act_feat[repo]:
this_repo_cluster_feat = []
for feat_name in repo_act_feat[repo][cluster]:
this_repo_cluster_feat += repo_act_feat[repo][cluster][
feat_name]
repo_act_feat[repo][cluster] = this_repo_cluster_feat
if sum(this_repo_cluster_feat) > max_sum_act_feat:
max_sum_act_feat = sum(this_repo_cluster_feat)
max_sum_act_feat_vec = this_repo_cluster_feat
print_and_log(logger, "max_sum_act_feat: {}".format(max_sum_act_feat))
print_and_log(logger,
"max_sum_act_feat_vec: {}".format(max_sum_act_feat_vec))
if config['use_user_profile_features']:
with open(config['user_profile_feat_path'], 'r') as f:
user_profile_feat = json.load(f)
if config['use_user_activity_features']:
with open(config['user_activity_feat_path'], 'r') as f:
user_act_feat = json.load(f)
if config['use_user_cluster_one_hot']:
with open(config['user_cluster_path'], 'r') as f:
user_clusters = json.load(f)
empty_no_event_user_cluster = max(user_clusters.values()) + 1
user_clusters[config['empty_user']] = empty_no_event_user_cluster
user_clusters['no_event_user'] = empty_no_event_user_cluster
if config['use_user_cluster_minmax']:
with open(config['user_cluster_path'], 'r') as f:
user_clusters = json.load(f)
max_cluster_id = max(user_clusters.values())
min_cluster_id = min(user_clusters.values())
empty_no_event_user_cluster = max_cluster_id + 1
max_cluster_id = empty_no_event_user_cluster
max_minus_min = max_cluster_id - min_cluster_id
if min_cluster_id != 0:
print("min cluster id is not 0! Need to examine code!")
pdb.set_trace()
user_clusters[config['empty_user']] = empty_no_event_user_cluster
user_clusters['no_event_user'] = empty_no_event_user_cluster
if config['use_cluster_profile_features']:
max_sum_profile_feat = 0
max_sum_profile_feat_vec = None
with open(config['cluster_profile_feat_path'], 'rb') as f:
cluster_profile_feat = pickle.load(f)
for cluster in cluster_profile_feat:
this_cluster_feat = []
for feat_name in cluster_profile_feat[cluster]:
# if feat_name == "geolocation" or feat_name == "user_type":
# continue
this_cluster_feat += cluster_profile_feat[cluster][feat_name]
cluster_profile_feat[cluster] = this_cluster_feat
if sum(this_cluster_feat) > max_sum_profile_feat:
max_sum_profile_feat = sum(this_cluster_feat)
max_sum_profile_feat_vec = this_cluster_feat
cluster_profile_feat[empty_no_event_user_cluster] = [
0
] * config['dim_cluster_profile_features']
print_and_log(logger,
"max_sum_profile_feat: {}".format(max_sum_profile_feat))
print_and_log(
logger,
"max_sum_profile_feat_vec: {}".format(max_sum_profile_feat_vec))
if config['use_cluster_activity_features']:
max_sum_act_feat = 0
max_sum_act_feat_vec = None
with open(config['cluster_activity_feat_path'], 'rb') as f:
cluster_act_feat = pickle.load(f)
for cluster in cluster_act_feat:
for repo in cluster_act_feat[cluster]:
this_cluster_repo_feat = []
for feat_name in cluster_act_feat[cluster][repo]:
this_cluster_repo_feat += cluster_act_feat[cluster][repo][
feat_name]
cluster_act_feat[cluster][repo] = this_cluster_repo_feat
if sum(this_cluster_repo_feat) > max_sum_act_feat:
max_sum_act_feat = sum(this_cluster_repo_feat)
max_sum_act_feat_vec = this_cluster_repo_feat
print_and_log(logger, "max_sum_act_feat: {}".format(max_sum_act_feat))
print_and_log(logger,
"max_sum_act_feat_vec: {}".format(max_sum_act_feat_vec))
print_and_log(logger, "x_dim: {}".format(x_dim))
user_has_no_cluster = set()
sample_id = 0
repo_y_et = []
repo_y_td = []
repo_y_uc = []
print_and_log(logger, "gather testing data...")
for repo in repo_list: # for each cascade chain
if (repo in config['repos_to_ignore']) or (repo
in github_event_repos_set):
# it is a event repo or repo should be ignore
continue
repo_X = []
repo2sampleid = []
one_chain_pd = load_jsongz_2_dataframe(
os.path.join(config['cascade_dir'], repo + '.json.gz'))
one_chain_pd = one_chain_pd.loc[
(one_chain_pd['nodeTime'] >= train_start)
& (one_chain_pd['nodeTime'] <= sim_end)]
one_chain_pd = one_chain_pd.sort_values(by=['nodeTime'])
one_chain_event = []
one_chain_time = []
one_chain_user = []
# padding event
for i in range(config['window_size']):
one_chain_event.append(config['empty_event_type'])
one_chain_time.append(config['empty_time_delay'])
one_chain_user.append(config['empty_user'])
# <soc>
one_chain_event.append(eventtype_2_id['<soc>'])
one_chain_time.append(config['empty_time_delay'])
one_chain_user.append(config['empty_user'])
# event sequence
one_chain_event += [
eventtype_2_id[event] for event in one_chain_pd['actionType']
]
one_chain_time += [time for time in one_chain_pd['nodeTime']]
one_chain_user += [user for user in one_chain_pd['nodeUserID']]
(one_chain_event_new, one_chain_time_new, one_chain_user_new) = \
insert_no_event_for_a_chain_new(config,
one_chain_event, one_chain_time,
one_chain_user, sim_end+1)
# if one_chain_event_new != one_chain_event:
# pdb.set_trace()
one_chain_event = one_chain_event_new
one_chain_time = one_chain_time_new
one_chain_user = one_chain_user_new
# calculate time delay sequence
one_chain_time_delay = []
for i in range(len(one_chain_time)):
if (one_chain_event[i] == config['empty_event_type']
or one_chain_event[i] == eventtype_2_id['<soc>']):
one_chain_time_delay.append(config['empty_time_delay'])
elif one_chain_event[i - 1] == eventtype_2_id['<soc>']:
one_chain_time_delay.append(config['empty_time_delay'])
else:
time_delay = get_time_delay(one_chain_time[i - 1],
one_chain_time[i], 'float')[1]
if config['time_delay_normalization_func'] is not None:
time_delay = time_delay_normalization(
time_delay, config['time_delay_normalization_func'])
one_chain_time_delay.append(time_delay)
# for each event sample in simulation period
for i in range(config['window_size'], len(one_chain_event)):
time_sample_outputevent = one_chain_time[i]
event_sample_outputevent = one_chain_event[i]
# if time_sample_outputevent in simulation period:
# add this sample to testset
if event_sample_outputevent == config['empty_event_type'] or (
event_sample_outputevent == eventtype_2_id['<soc>']):
continue
if one_chain_event[i - 1] == eventtype_2_id['<soc>']:
continue
if not ((time_sample_outputevent >= sim_start) and
(time_sample_outputevent <= sim_end)):
continue
input_event_type = \
one_chain_event[i-config['window_size']:i]
# input_time = one_chain_time[i-config['window_size']:i]
input_time_delay = \
one_chain_time_delay[i-config['window_size']:i]
input_user = one_chain_user[i - config['window_size']:i]
input_cluster = []
for user in input_user:
try:
input_cluster.append(user_clusters[user])
except KeyError:
user_has_no_cluster.add(user)
input_cluster.append(user_clusters['no_event_user'])
output_event_type = one_chain_event[i]
output_time_delay = one_chain_time_delay[i]
output_user = one_chain_user[i]
try:
output_cluster = user_clusters[output_user]
except KeyError:
user_has_no_cluster.add(output_user)
output_cluster = user_clusters['no_event_user']
# initialize input vector, and output vector for this sample
x_vec = []
# load repo embeding vector
if config['use_repo_embedding']:
try:
embed_vec = np.array(embed_vec_all[repo[:22] + '/' +
repo[23:]])
except KeyError:
print_and_log(
logger, "Could not find "
"embedding vector for {}!".format(repo[:22] + '/' +
repo[23:]))
pdb.set_trace()
# input feature vector
for j in range(config['window_size']): # for each event node
x_j = []
if config['use_repo_embedding']:
x_j += list(embed_vec)
if config['use_repo_profile_features']:
try:
x_j += repo_profile_feat[repo]
except KeyError:
x_j += [0] * config['dim_repo_profile_features']
if config['use_repo_idx_features']:
try:
x_j += load_repo_train_vali[repo]
except KeyError:
x_j += [0]
if config['use_repo_activity_features']:
if input_cluster[j] == empty_no_event_user_cluster:
x_j += [0] * config['dim_repo_activity_features']
else:
try:
repo_thiscluster_act_feat = repo_act_feat[repo][
input_cluster[j]]
repo_allcluster_act_feat = repo_act_feat[repo][
'all_cluster']
x_j += repo_thiscluster_act_feat
x_j += repo_allcluster_act_feat
except KeyError:
x_j += [0] * config['dim_repo_activity_features']
if config['use_user_profile_features']:
if input_user[j] == config['empty_user'] or (
input_user[j] == 'no_event_user'):
x_j += [0] * config['dim_user_profile_features']
else:
try:
x_j += user_profile_feat[input_user[j]]
except KeyError:
x_j += [0] * config['dim_user_profile_features']
if config['use_user_activity_features']:
if input_user[j] == config['empty_user'] or (
input_user[j] == 'no_event_user'):
x_j += [0] * config['dim_user_activity_features']
else:
try:
thisrepo_feat = \
user_act_feat[input_user[j]][repo]
except KeyError:
# this user-repo no event in training period
thisrepo_feat = \
[0] * int(config[
'dim_user_activity_features']/2)
allrepo_feat = \
user_act_feat[input_user[j]]['all']
x_j += thisrepo_feat + allrepo_feat
if config['use_event_type_one_hot']:
event_type_one_hot = \
[0] * len(config['eventtype_2_id'])
if input_event_type[j] != config['empty_event_type']:
event_type_one_hot[input_event_type[j] - 1] = 1
x_j += event_type_one_hot
if config['use_time_delay_features']:
x_j += [input_time_delay[j]]
if config['use_user_cluster_one_hot']:
user_cluster_one_hot = \
[0] * config['dim_user_cluster_one_hot']
user_cluster_one_hot[input_cluster[j]] = 1
x_j += user_cluster_one_hot
if config['use_user_cluster_minmax']:
use_user_cluster_minmax = (input_cluster[j] /
max_minus_min)
x_j += [use_user_cluster_minmax]
if config['use_cluster_profile_features']:
this_cluster_profile_feat = cluster_profile_feat[
input_cluster[j]]
x_j += this_cluster_profile_feat
if config['use_cluster_activity_features']:
if input_cluster[j] == empty_no_event_user_cluster:
x_j += [0] * config['dim_cluster_activity_features']
else:
try:
cluster_thisrepo_act_feat = cluster_act_feat[
input_cluster[j]][repo]
cluster_allrepo_act_feat = cluster_act_feat[
input_cluster[j]]['all_repo']
x_j += cluster_thisrepo_act_feat
x_j += cluster_allrepo_act_feat
except KeyError:
x_j += [0
] * config['dim_cluster_activity_features']
if len(x_j) != x_dim:
print("len(x_j) != x_dim")
pdb.set_trace()
x_vec.append(x_j)
if len(x_vec) != config['window_size']:
print("len(x_vec) != config['window_size']")
pdb.set_trace()
repo_X.append(x_vec)
# output vec
repo_y_et.append([output_event_type])
repo_y_td.append([output_time_delay])
repo_y_uc.append([output_cluster])
repo2sampleid.append(sample_id)
sample_id += 1
# pdb.set_trace()
# finish gathering test data for this repo
testset['X_test'] += repo_X
testset['repo2sampleid'].append(repo2sampleid)
# pdb.set_trace()
testset['X_test'] = np.array(testset['X_test'])
testset['Y_test_et'] = np.array(repo_y_et)
testset['Y_test_td'] = np.array(repo_y_td)
testset['Y_test_uc'] = np.array(repo_y_uc)
print_and_log(logger, "X_test.shape: {}".format(testset['X_test'].shape))
print_and_log(logger,
"Y_test_et.shape: {}".format(testset['Y_test_et'].shape))
print_and_log(logger,
"Y_test_td.shape: {}".format(testset['Y_test_td'].shape))
print_and_log(logger,
"Y_test_uc.shape: {}".format(testset['Y_test_uc'].shape))
print_and_log(logger,
"number of chains for testing: {}".format(len(repo_list)))
print_and_log(
logger, "could not find cluster for {} users.".format(
len(user_has_no_cluster)))
if len(user_has_no_cluster) > 0:
with open(
os.path.join(config['exp_save_dir'], "dataset",
'user_has_no_cluster_given_gt.pickle'),
'wb') as f:
pickle.dump(user_has_no_cluster, f)
print_and_log(
logger,
"the number of testing samples: {}".format(len(testset['X_test'])))
with open(
os.path.join(config['exp_save_dir'], "dataset",
'testset_given_gt.pickle'), 'wb') as f:
pickle.dump(testset, f)
print_and_log(
logger,
"{} took {} min".format("testset creation given gt",
(time.time() - create_dataset_start) / 60))
# pdb.set_trace()
return testset
def testset_creation(config, GT_avail=True):
create_dataset_start = time.time()
logger = set_logger(
os.path.join(
config['exp_save_dir'], 'testset_creation_' +
dt.now().strftime("%Y-%m-%dT%H-%M-%SZ") + '.log'))
print_and_log(logger, '{}'.format('testset creation...'))
if not os.path.exists(os.path.join(config['exp_save_dir'], "dataset")):
os.makedirs(os.path.join(config['exp_save_dir'], "dataset"))
testset = dict()
testset['X_test'] = []
testset['repo'] = []
testset['input_last_event_time'] = []
testset['gt_event_type'] = None
testset['gt_event_id'] = None
testset['gt_time_delay'] = None
testset['gt_user'] = None
testset['gt_user_cluster'] = None
X_test = []
eventtype_2_id = config['eventtype_2_id']
id_2_eventtype = dict(zip(eventtype_2_id.values(), eventtype_2_id.keys()))
sim_start = utc_timestamp(config['sim_period']['start'])
sim_end = utc_timestamp(config['sim_period']['end'])
with open(config['unique_repo_train_vali_path'], 'r') as f:
unique_repo_train_vali = json.load(f)
repo_list = list(unique_repo_train_vali.keys())
with open(config['github_event_repos_path'], 'rb') as f:
github_event_repos = pickle.load(f)
github_event_repos_set = set(
[repo[:22] + '-' + repo[23:] for repo in github_event_repos])
with open(config['user_cluster_path'], 'r') as f:
user_clusters = json.load(f)
empty_no_event_user_cluster = max(user_clusters.values()) + 1
user_clusters[config['empty_user']] = empty_no_event_user_cluster
user_clusters['no_event_user'] = empty_no_event_user_cluster
unique_repo_test = dict()
unique_user_test = dict()
user_has_no_cluster = set()
# gather simulation sample input
print_and_log(logger, "gather simulation sample input...")
for repo in repo_list: # for each cascade chain
if (repo in config['repos_to_ignore']) or (repo
in github_event_repos_set):
# it is a event repo or repo should be ignore
continue
one_chain_pd = load_jsongz_2_dataframe(
os.path.join(config['cascade_dir'], repo + '.json.gz'))
# get all events before sim
one_chain_pd = one_chain_pd.loc[one_chain_pd['nodeTime'] < sim_start]
one_chain_pd = one_chain_pd.sort_values(by=['nodeTime'])
one_chain_event = []
one_chain_time = []
one_chain_user = []
# padding event
for i in range(config['window_size']):
one_chain_event.append(config['empty_event_type'])
one_chain_time.append(config['empty_time_delay'])
one_chain_user.append(config['empty_user'])
# <soc>
one_chain_event.append(eventtype_2_id['<soc>'])
one_chain_time.append(config['empty_time_delay'])
one_chain_user.append(config['empty_user'])
# event sequence
one_chain_event += [
eventtype_2_id[event] for event in one_chain_pd['actionType']
]
one_chain_time += [time for time in one_chain_pd['nodeTime']]
one_chain_user += [user for user in one_chain_pd['nodeUserID']]
(one_chain_event_new, one_chain_time_new,
one_chain_user_new) = \
insert_no_event_for_a_chain_new(config, one_chain_event,
one_chain_time, one_chain_user,
sim_start)
# if one_chain_event_new != one_chain_event:
# pdb.set_trace()
one_chain_event = one_chain_event_new
one_chain_time = one_chain_time_new
one_chain_user = one_chain_user_new
# calculate time delay sequence
one_chain_time_delay = []
for i in range(len(one_chain_time)):
if (one_chain_event[i] == config['empty_event_type']
or one_chain_event[i] == eventtype_2_id['<soc>']):
one_chain_time_delay.append(config['empty_time_delay'])
elif one_chain_event[i - 1] == eventtype_2_id['<soc>']:
one_chain_time_delay.append(config['empty_time_delay'])
else:
time_delay = get_time_delay(one_chain_time[i - 1],
one_chain_time[i], 'float')[1]
if config['time_delay_normalization_func'] is not None:
time_delay = time_delay_normalization(
time_delay, config['time_delay_normalization_func'])
one_chain_time_delay.append(time_delay)
# input_last_event_time
testset['input_last_event_time'].append(one_chain_time[-1])
testset['repo'].append(repo)
input_event_type = one_chain_event[-config['window_size']:]
input_time_delay = one_chain_time_delay[-config['window_size']:]
input_user = one_chain_user[i - config['window_size']:i]
input_cluster = []
for user in input_user:
try:
input_cluster.append(user_clusters[user])
except KeyError:
user_has_no_cluster.add(user)
input_cluster.append(user_clusters['no_event_user'])
# input feature vector
if len(input_event_type) < config['window_size']:
print("len(input_event_type) < config['window_size']")
pdb.set_trace()
# initialize input for this sample
hightd_normalized = time_delay_normalization(
720, config['time_delay_normalization_func'])
lowtd_normalized = time_delay_normalization(
0.001, config['time_delay_normalization_func'])
td_normalized2 = time_delay_normalization(
0.1, config['time_delay_normalization_func'])
last_et = config['eventtype_2_id']['no_event_for_1month']
last_td = hightd_normalized
last_uc = user_clusters['no_event_user']
x_vec = [last_et, last_td, last_uc]
# if repo == "75Q5j4D5taKq5AlL--ZIFg-WKcY0zAnxfYZx6laSNH2UA":
# pdb.set_trace()
# if repo == "2YzVcEU5XvJXobTU6swknA-uD2kFk9smPOUWFVcpo1bXg":
# pdb.set_trace()
X_test.append(x_vec)
testset['X_test'] = X_test
print_and_log(logger, "X_test length: {}".format(len(testset['X_test'])))
# gather simulation sample output
if GT_avail:
print_and_log(
logger, "ground truth available. \n"
"gather simulation sample output...")
testset['gt_event_type'] = []
testset['gt_time_delay'] = []
testset['gt_event_id'] = []
testset['gt_user'] = []
testset['gt_user_cluster'] = []
# for each gathered simulation sample input
for i in range(len(testset['repo'])):
repo_this_sample = testset['repo'][i]
one_chain_pd = load_jsongz_2_dataframe(
os.path.join(config['cascade_dir'],
repo_this_sample + '.json.gz'))
# get events during sim
one_chain_pd = one_chain_pd.loc[
(one_chain_pd['nodeTime'] >= sim_start)
& (one_chain_pd['nodeTime'] <= sim_end)]
one_chain_pd = one_chain_pd.sort_values(by=['nodeTime'])
one_chain_event_id = []
one_chain_time = []
one_chain_user = []
unique_repo_test[repo_this_sample] = []
if len(one_chain_pd) == 0:
# this repo has no events in testing period
# need to insert no event till end of testing
input_last_event_time_this_chain = testset[
'input_last_event_time'][i]
(one_chain_event_new, one_chain_time_new, one_chain_user_new
) = insert_no_event_for_a_GTchain_who_has_no_event_at_all(
config, one_chain_event_id, one_chain_time,
one_chain_user, input_last_event_time_this_chain, sim_end)
else:
# this sample chain has events in testing period
for user in one_chain_pd['nodeUserID']:
unique_user_test[user] = []
one_chain_event_id += [
eventtype_2_id[event]
for event in one_chain_pd['actionType']
]
one_chain_time += [time for time in one_chain_pd['nodeTime']]
one_chain_user += [user for user in one_chain_pd['nodeUserID']]
input_last_event_time_this_chain = testset[
'input_last_event_time'][i]
if min(one_chain_time) < sim_start:
print("min(one_chain_time) < sim_start")
pdb.set_trace()
(one_chain_event_new, one_chain_time_new,
one_chain_user_new) = \
insert_no_event_for_a_sim_GTchain(
config,
one_chain_event_id, one_chain_time,
one_chain_user,
input_last_event_time_this_chain,
sim_end
)
# if one_chain_event_new != one_chain_event:
# pdb.set_trace()
one_chain_event_id = one_chain_event_new
one_chain_time = one_chain_time_new
one_chain_user = one_chain_user_new
one_chain_event_type = [
id_2_eventtype[event] for event in one_chain_event_id
]
one_chain_user_cluster = []
# no mather one-hot or normalized, here need true cluster
for user in one_chain_user:
try:
one_chain_user_cluster.append(user_clusters[user])
except KeyError:
user_has_no_cluster.add(user)
one_chain_user_cluster.append(
user_clusters['no_event_user'])
# calculate time delay sequence
one_chain_time_delay = []
if len(one_chain_time) > 0:
time_delay = get_time_delay(input_last_event_time_this_chain,
one_chain_time[0], 'float')[1]
# NO NEED TO DO normalization FOR GT !!!
# if config['time_delay_normalization_func'] is not None:
# time_delay = time_delay_normalization(
# time_delay,
# config['time_delay_normalization_func'])
one_chain_time_delay.append(time_delay)
for j in range(1, len(one_chain_time)):
time_delay = get_time_delay(one_chain_time[j - 1],
one_chain_time[j], 'float')[1]
# NO NEED TO DO normalization FOR GT !!!
# if config[
# 'time_delay_normalization_func'] is not None:
# time_delay = time_delay_normalization(
# time_delay,
# config['time_delay_normalization_func'])
one_chain_time_delay.append(time_delay)
testset['gt_event_id'].append(one_chain_event_id)
testset['gt_event_type'].append(one_chain_event_type)
testset['gt_time_delay'].append(one_chain_time_delay)
testset['gt_user'].append(one_chain_user)
testset['gt_user_cluster'].append(one_chain_user_cluster)
# pdb.set_trace()
print_and_log(
logger, "could not find cluster for {} users.".format(
len(user_has_no_cluster)))
# check num of 0.0 hours in the ground truth output
gt_zero_hours_count = 0
gt_hours_count = 0
for i in range(len(testset['gt_time_delay'])):
for j in range(len(testset['gt_time_delay'][i])):
time_delay_hour = testset['gt_time_delay'][i][j]
gt_hours_count += 1
if time_delay_hour == 0:
gt_zero_hours_count += 1
print_and_log(
logger, "Out of {} ground truth time delay values that the model "
"needs to predict, {} of them are 0.0 time delay hour.".format(
gt_hours_count, round(gt_zero_hours_count / gt_hours_count, 2)))
# save testset
print_and_log(logger, "save testset ...")
testset_save_path = os.path.join(config['exp_save_dir'], "dataset",
'testset.pickle')
with open(testset_save_path, 'wb') as handle:
pickle.dump(testset, handle, protocol=pickle.HIGHEST_PROTOCOL)
print_and_log(
logger, "testset.pickle save in {}".format(
os.path.join(config['exp_save_dir'], "dataset")))
with open(
os.path.join(config['exp_save_dir'], "dataset",
'unique_repo_test.json'), 'w') as f:
json.dump(unique_repo_test, f)
with open(
os.path.join(config['exp_save_dir'], "dataset",
'unique_user_test.json'), 'w') as f:
json.dump(unique_user_test, f)
print_and_log(
logger, "the number of unique repos in "
"testing samples: {}".format(len(unique_repo_test)))
print_and_log(
logger, "the number of unique users in "
"testing samples: {}".format(len(unique_user_test)))
print_and_log(
logger, "{} took {}".format("testset creation",
time.time() - create_dataset_start))
# pdb.set_trace()
return testset