outliers.py

# Use StreamAD library: https://github.com/Fengrui-Liu/StreamAD
# The first three algorithms have lower scores for outliers, while the others have higher scores for outliers. Thus, the AUC ROC score x should be 1-x for the first three algorithms.
# To test these stream outlier dectors, just run `python outliers.py`.

import numpy as np
import logging
import datetime
from pipeline import *
from streamad.model import xStreamDetector, RShashDetector, HSTreeDetector, LodaDetector, RrcfDetector
from sklearn.metrics import roc_auc_score

def outlier_detector_marker(data):
    
    seed=42
    
    model_dict = {
        'ECOD':PYOD,
        'IForest':PYOD,
    }

    anomaly_list = []
    
    for name, clf in model_dict.items():
        #logger.info(name)
        clf = clf(seed=seed, model_name=name)
        clf = clf.fit(data, [])
        # output predicted anomaly score on testing set
        score = clf.predict_score(data)
        #logger.info(score)
        t = score.mean()+2*score.std()
        #logger.info(t)
        anomaly_list.append(np.where(score>t,1,0))
    
    anomaly_index = anomaly_list[0]
    for i in range(1,len(anomaly_list)):
        anomaly_index = anomaly_index * anomaly_list[i]

    # logger.info(sum(anomaly_index)/len(anomaly_index))        
    
    return anomaly_index
    # return outlier_stats_each_window, outlier_stats_overall



if __name__ == "__main__":
    mkdirs("logs/")

    for handler in logging.root.handlers[:]:
        logging.root.removeHandler(handler)

    log_file_name = 'logs/experiment_log-%s.log' % (datetime.datetime.now().strftime("%Y-%m-%d-%H:%M-%S"))

    logging.basicConfig(
        filename=log_file_name,
        format='%(asctime)s %(levelname)-8s %(message)s',
        datefmt='%m-%d %H:%M', level=logging.INFO, filemode='w')

    logger = logging.getLogger()
    logger.setLevel(logging.INFO)

    dataset_prefix_list = ['dataset_experiment_info/room_occupancy', 'dataset_experiment_info/electricity_prices', 'dataset_experiment_info/insects/incremental_reoccurring_balanced', 'dataset_experiment_info/beijing_multisite/shunyi', 'dataset_experiment_info/tetouan']
    done = []

    
    for dataset_path_prefix in dataset_prefix_list:      
        if dataset_path_prefix in done:
            continue
        
        streamad_models = [xStreamDetector(depth=10), RShashDetector(components_num=10), HSTreeDetector(), LodaDetector(), RrcfDetector()]

        logger.info(dataset_path_prefix)
        try:
            data_path, schema_path, task = schema_parser(dataset_path_prefix)
        except:
            continue
        
        logger.info("start pre-processing")
        
        target_data_nonnull, data_before_onehot, data_onehot_nonnull, original_columns, window_size, row_count, original_column_count, new_columns, new_column_count = data_preprocessing(dataset_path_prefix, data_path, schema_path, task, logger)
        logger.info("preprocessing done")

        outlier_labels = outlier_detector_marker(data_onehot_nonnull)

        data_onehot_nonnull = data_onehot_nonnull.to_numpy()

        for model in streamad_models:
            scores = []
            for i in range(len(data_onehot_nonnull)):
                score = model.fit_score(data_onehot_nonnull[i])
                scores.append(score)
            # logger.info(scores)
            auc = roc_auc_score(outlier_labels[100:], scores[100:])
            logger.info(auc)