hotsax

HOT-SAX

Citekey	KeoghEtAl2005HOT
Source Code	https://github.com/seninp/saxpy
Learning type	unsupervised
Input dimensionality	univariate

Notes

We use the source code from the Github repository instead of the published library on pypi because it contains runtime improvements and edge case fixes.

HOT-SAX outputs the distance of discords to their nearest non-self match. Therefore, the results require post-processing.

U can use the following code snippet for the post-processing step in TimeEval (default parameters directly filled in from the source code):

import pandas as pd
import numpy as np

from scipy.sparse import csc_matrix, hstack

from timeeval.utils.window import ReverseWindowing
from timeeval import AlgorithmParameter

# post-processing for HOT-SAX
def post_hotsax(algorithm_parameter: AlgorithmParameter, args: dict) -> np.ndarray:
    window_size = args.get("hyper_params", {}).get("anomaly_window_size", 100)
    if isinstance(algorithm_parameter, np.ndarray):
        results = pd.DataFrame(algorithm_parameter)
    else:
        results = pd.read_csv(algorithm_parameter)
    results.columns = ["score"]
    anomalies = results[results["score"] > .0]

    # use scipy sparse matrix to save memory
    matrix = csc_matrix((len(results), 1), dtype=np.float64)
    counts = np.zeros(len(results))
    for i, row in anomalies.iterrows():
        idx = int(row.name)
        tmp = np.zeros(len(results))
        tmp[idx:idx + window_size] = np.repeat([row["score"]], repeats=window_size)
        tmp = tmp.reshape(-1, 1)
        matrix = hstack([matrix, tmp])
        counts[idx:idx + window_size] += 1
    sums = matrix.sum(axis=1)
    counts = counts.reshape(-1, 1)
    scores = np.zeros_like(sums)
    np.divide(sums, counts, out=scores, where=counts != 0)
    # returns the completely flattened array (from `[[1.2], [2.3]]` to `[1.2, 2.3]`)
    return scores.A1  # type: ignore

Citation format

Senin, P., Lin, J., Wang, X., Oates, T., Gandhi, S., Boedihardjo, A.P., Chen, C., Frankenstein, S., Lerner, M., GrammarViz 2.0: a tool for grammar-based pattern discovery in time series, ECML/PKDD Conference, 2014.