BENCHMARK.md

Benchmark

This document explains the benchmarking procedure we develop in Orion in order to evaluate how good a pipeline is detecting anomalies.

Evaluating the Pipelines

Using the Evaluation sub-package, we can compute a score given a set of known anomalies and another one of detected anomalies.

We can evaluate the overall performance of our pipelines in order to know which one has the best performance.

For this we:

1. Use each pipeline to detect anomalies on all datasets and their signals.
2. Retreive the list of known anomalies for each of these signals.
3. Compute the scores for each signal using multiple metrics (e.g. accuraccy and f1).
4. Average the score obtained for each metric and pipeline accross all the signals.
5. Finally, we rank our pipelines sorting them by one of the computed scores.

Benchmark Function

Pipelines

The first item required to benchmark is the set of pipelines you want to evaluate, this can be done by supplying a list of pipelines or a dictionary of pipelines.

pipelines = [ 
    'arima',
    'lstm_dynamic_threshold'
]

You can use pipelines from the suite of existing pipelines in Orion or, alternatively, you can use your own implemented pipeline.

Datasets

An important item in benchmarking is the requirement to know the ground truth anomalies for proper comparison. An example of such dataset is the demo signals and a list of their known anomalies, obtained from the telemanom repository. Using these signals, we can compute how similar the anomalies detected by the pipeline are with the prevoiusly known anomalies.

Similarly, we define the datasets as either a single dataset compose of a list of signals, or a dictionary of datasets.

signals = [
    'S-1',
    'P-1'
]

# or defined by a dataset name
datasets = {
    'demo': signals
}

Hyperparameters

In most cases, pipelines need to be modified based on the dataset currently being evaluated. To provide this functionality, you can use hyperparameters, which are nested dictionaries, to change a particular hyperparameter setting within the pipeline. For example, to consider changing the number of epochs within lstm_dynamic_threshold pipeline for the demo dataset, we use:

hyperparameters = {
    'demo': {
        'lstm_dynamic_threshold': {
            'keras.Sequential.LSTMTimeSeriesRegressor#1': {
                'epochs': 5
            }
        }
    }
}

Arguments

To use orion.benchmark.benchmark function, it expects the following inputs:

• pipelines (dict or list): dictionary with pipeline names as keys and their JSON paths as values. If a list is given, it should be of JSON paths, and the paths themselves will be used as names. If not given, all verified pipelines will be used for evaluation.
• datasets (dict or list): dictionary of dataset name as keys and list of signals as values. If a list is given then it will be under a generic name dataset. If not given, all benchmark datasets will be used used.
• hyperparameters (dict or list): dictionary with pipeline names as keys and their hyperparameter JSON paths or dictionaries as values. If a list is given, it should be of corresponding order to pipelines.
• metrics (dict or list): dictionary with metric names as keys and scoring functions as values. If a list is given, it should be of scoring functions, and they __name__ value will be used as the metric name. If not given, all the available metrics will be used.
• rank (str): Sort and rank the pipelines based on the given metric. If not given, rank using the first metric.
• test_split (bool or float): if bool, it indicates whether to use the prespecified train-test split. If float, then it should be between 0.0 and 1.0 and represent the proportion of the signal to include in the test split. If not given, use False.
• detrend (bool): Whether to use scipy.detrend. If not given, use False.
• iterations (int): Number of iterations to perform over each signal and pipeline. Defaults to 1.
• workers (int or str): If workers is given as an integer value other than 0 or 1, a multiprocessing Pool is used to distribute the computation across the indicated number of workers. If the string dask is given, the computation is distributed using dask. In this case, setting up the dask cluster and client is expected to be handled outside of this function.
• show_progress (bool): Whether to use tqdm to keep track of the progress. Defaults to True.
• cache_dir (str): If a cache_dir is given, intermediate results are stored in the indicated directory as CSV files as they get computted. This allows inspecting results while the benchmark is still running and also recovering results in case the process does not finish properly. Defaults to None.
• output_path (str): Location to save the intermediatry results. If not given, intermediatry results will not be saved.
• pipeline_dir (str): If a pipeline_dir is given, pipelines will get dumped in the specificed directory as pickle files. Defaults to None.

And returns a pandas.DataFrame which contains the scores obtained with each scoring function accross all the signals for each pipeline.

This is an example of how to call this function:

from orion.benchmark import benchmark

pipelines = [
    'arima',
    'lstm_dynamic_threshold'
]

signals = ['S-1', 'P-1']

metrics = ['f1', 'accuracy', 'recall', 'precision']

scores = benchmark(pipelines=pipelines, datasets=signals, metrics=metrics, rank='f1')

⚠️ Benchmarking might require a long time to compute.

The output of the benchmark will be a pandas.DataFrame containing the results obtained by each pipeline on evaluating each signal.

                 pipeline  rank  accuracy  elapsed     f1         precision   recall   
0  lstm_dynamic_threshold     1  0.986993  915.958132  0.846868   0.879518     0.816555
1                   arima     2  0.962160  319.968949  0.382637   0.680000     0.266219 

Releases

In every release, we run Orion benchmark. We maintain an up-to-date leaderboard with the current scoring of the verified pipelines according to the benchmarking procedure explained above.

We run the benchmark on 11 datasets with their known grounth truth. We record the score of the pipelines on each dataset. Results obtained during the benchmarking process as well as previous benchmarks can be found within benchmark/results folder as CSV files. In addition, you can find it in the details Google Sheets document.

Leaderboard

We summarize the results in the leaderboard table. We showcase the number of wins each pipeline has over the ARIMA pipeline.

The summarized results can also be browsed in the following summary Google Sheets document.