Time Smash

Time series clustering and classification suite using notions of Universal similarity among data streams, especially without a priori knowledge about the "correct" features to use for time series data.

• Featurization algorithms: SymbolicDerivative, InferredHMMLikelihood, Csmash
• Distance measure: LikelihoodDistance

Example publications

• Huang, Yi, Victor Rotaru, and Ishanu Chattopadhyay. "Sequence likelihood divergence for fast time series comparison." Knowledge and Information Systems 65, no. 7 (2023): 3079-3098. https://link.springer.com/article/10.1007/s10115-023-01855-0

• Chattopadhyay, Ishanu, and Hod Lipson. "Data smashing: uncovering lurking order in data." Journal of The Royal Society Interface 11, no. 101 (2014): 20140826. https://royalsocietypublishing.org/doi/10.1098/rsif.2014.0826

• Timesmash: Process-Aware Fast Time Series Clustering and Classification https://easychair.org/publications/preprint/qpVv

For questions or suggestions contact:research@paraknowledge.ai

Usage examples

SymbolicDerivative

from timesmash import SymbolicDerivative
from sklearn.ensemble import RandomForestClassifier

train = [[1, 0, 1, 0, 1, 0], [1, 1, 0, 1, 1, 0]]
train_label = [[0], [1]]
test = [[0, 1, 1, 0, 1, 1]]
train_features, test_features = SymbolicDerivative().fit_transform(
    train=train, test=test, label=train_label
)
clf = RandomForestClassifier().fit(train_features, train_label)
label = clf.predict(test_features)
print("Predicted label: ", label)

LikelihoodDistance

from timesmash import LikelihoodDistance
from sklearn.cluster import KMeans
train = [[1, 0, 1.1, 0, 11.2, 0], [1, 1, 0, 1, 1, 0], [0, 0.9, 0, 1, 0, 1], [0, 1, 1, 0, 1, 1]]
dist_calc = LikelihoodDistance().fit(train)
dist = dist_calc.produce()
from sklearn.cluster import KMeans
clusters = KMeans(n_clusters = 2).fit(dist).labels_
print("Clusters: " clusters)

InferredHMMLikelihood

from timesmash import InferredHMMLikelihood
from sklearn.ensemble import RandomForestClassifier

train = [[1, 0, 1, 0, 1, 0], [1, 1, 0, 1, 1, 0]]
train_label = [[0], [1]]
test = [[0, 1, 1, 0, 1, 1]]
train_features, test_features = InferredHMMLikelihood().fit_transform(
    train=train, test=test, label=train_label
)
clf = RandomForestClassifier().fit(train_features, train_label)
label = clf.predict(test_features)
print("Predicted label: ", label)

ClusteredHMMClassifier:

from timesmash import Quantizer, InferredHMMLikelihood, LikelihoodDistance
from sklearn.cluster import KMeans
from sklearn.ensemble import RandomForestClassifier
import pandas as pd

train = pd.DataFrame(
    [[1, 0, 1, 0, 1, 0], [1, 1, 0, 1, 1, 0], [1, 0, 1, 0, 1, 0], [1, 1, 0, 1, 1, 0]]
)
train_label = pd.DataFrame([[0], [1], [0], [1]])
test = pd.DataFrame([[0, 1, 1, 0, 1, 1]])

qtz = Quantizer().fit(train, label=train_label)
new_labels = train_label.copy()
for label, dataframe in train_label.groupby(train_label.columns[0]):
    dist = LikelihoodDistance(quantizer=qtz).fit(train.loc[dataframe.index]).produce()
    sub_labels = KMeans(n_clusters=2, random_state=0).fit(dist).labels_
    new_label_names = [str(label) + "_" + str(i) for i in sub_labels]
    new_labels.loc[dataframe.index, train_label.columns[0]] = new_label_names

featurizer = InferredHMMLikelihood(quantizer=qtz, epsilon=0.01)
train_features, test_features = featurizer.fit_transform(
    train=train, test=test, label=new_labels
)

clf = RandomForestClassifier().fit(train_features, train_label)
print("Predicted label: ", clf.predict(test_features))

XHMMFeatures for anomaly detection:

import pandas as pd
from timesmash import XHMMFeatures
from sklearn.neighbors import LocalOutlierFactor

channel1_train = pd.DataFrame([[0,1,0,1,0,1,0,1,0,1],[1,0,1,0,1,0,1,0,1,0]], index=['person_1', 'person_2'])
channel2_train = pd.DataFrame([[0,1,0,1,0,1,0,1,0,1],[1,0,1,0,1,0,1,0,1,0]], index=['person_1', 'person_2'])
labels = pd.DataFrame([1,1], index=['person_1', 'person_2'])
    
alg = XHMMFeatures(n_quantizations=1)
features_train = alg.fit_transform([channel1_train,channel2_train], labels)
    
clf = LocalOutlierFactor(novelty=True)  
clf.fit(features_train)
        
channel1_test = pd.DataFrame([[0,1,0,1,0,1,0,1,0,1],[1,0,1,0,1,0,1,0,1]], index=['person_test_1', 'person_test_2'])
channel2_test= pd.DataFrame([[0,1,0,1,0,1,0,1,0,1],[0,1,0,1,0,1,0,1,0]], index=['person_test_1', 'person_test_2'])

features_test = alg.transform([channel1_test,channel2_test])
print(clf.predict(features_test))

XHMMFeatures for classification:

import pandas as pd
from timesmash import XHMMFeatures
from sklearn.ensemble import RandomForestClassifier

d1_train = pd.DataFrame([[0,1,0,1,0,1,0,1,0,1],[1,0,1,0,1,0,1,0,1,0]], index=['person_1', 'person_2'])
d2_train = pd.DataFrame([[1,0,1,0,1,0,1,0,1,0],[1,0,1,0,1,0,1,0,1,0]], index=['person_1', 'person_2'])
labels = pd.DataFrame([0,1], index=['person_1', 'person_2'])
    
alg = XHMMFeatures(n_quantizations=1)
features_train = alg.fit_transform([d1_train,d2_train], labels)
    
clf = RandomForestClassifier()  
clf.fit(features_train, labels)
        
d1_test = pd.DataFrame([[1,0,1,0,1,0,1,0,1]], index=['person_test'])
d2_test= pd.DataFrame([[0,1,0,1,0,1,0,1,0]], index=['person_test'])

features_test = alg.transform([d1_test,d2_test])
    
print(clf.predict(features_test))

XHMMClustering for multichannel clustering:

import pandas as pd
from timesmash import XHMMClustering

channel1 = pd.DataFrame(
    [
        [0, 1, 0, 1, 0, 1, 0, 1, 0, 1],
        [1, 0, 1, 0, 1, 0, 1, 0, 1, 0],
        [1, 0, 1, 0, 1, 0, 1, 0, 1, 0],
        [0, 1, 0, 1, 0, 1, 0, 1, 0, 1],
    ],
    index=["person_1", "person_2", "person_3", "person_4"],
)
channel2 = pd.DataFrame(
    [
        [1, 0, 1, 0, 1, 0, 1, 0, 1, 0],
        [1, 0, 1, 0, 1, 0, 1, 0, 1, 0],
        [0, 1, 0, 1, 0, 1, 0, 1, 0, 1],
        [0, 1, 0, 1, 0, 1, 0, 1, 0, 1],
    ],
    index=["person_1", "person_2", "person_3", "person_4"],
)
alg = XHMMClustering(n_quantizations=1).fit(
    [channel1, channel2]
)
clusters = alg.labels_
print(clusters)