Time-Series Classification

Overview

Time-series classification assigns categorical labels to entire sequences based on their temporal patterns. Samay models like MOMENT excel at classification by learning rich embeddings from time-series data.

Models Supporting Classification

Model	Zero-Shot	Fine-Tuning	Approach
MOMENT	✅ (via SVM)	✅	Embedding-based

Step-by-Step Workflow

Load model for classification

Initialize MOMENT with classification configuration:

from samay.model import MomentModel

repo = "AutonLab/MOMENT-1-large"
config = {
    "task_name": "classification",
    "n_channels": 1,
    "num_class": 5  # Number of classes in your dataset
}
mmt = MomentModel(config=config, repo=repo)

Prepare classification dataset

Load training and test data:

from samay.dataset import MomentDataset

train_dataset = MomentDataset(
    name="ecg5000",
    path="data/ECG5000_TRAIN.csv",
    batchsize=64,
    mode="train",
    task_name="classification",
)

test_dataset = MomentDataset(
    name="ecg5000",
    path="data/ECG5000_TEST.csv",
    batchsize=64,
    mode="test",
    task_name="classification",
)

Zero-shot classification (SVM approach)

Extract embeddings and train a simple classifier:

from samay.models.moment.momentfm.models.statistical_classifiers import fit_svm

# Extract embeddings
train_accuracy, train_embeddings, train_labels = mmt.evaluate(
    train_dataset, task_name="classification"
)
test_accuracy, test_embeddings, test_labels = mmt.evaluate(
    test_dataset, task_name="classification"
)

print(train_embeddings.shape, train_labels.shape)
# (500, 1024) (500,)

# Train SVM on embeddings
clf = fit_svm(features=train_embeddings, y=train_labels)

# Evaluate
train_accuracy = clf.score(train_embeddings, train_labels)
test_accuracy = clf.score(test_embeddings, test_labels)

print(f"Train accuracy: {train_accuracy:.2f}")
print(f"Test accuracy: {test_accuracy:.2f}")
# Train accuracy: 1.00
# Test accuracy: 0.93

Fine-tune for better performance

Train the classification head end-to-end:

finetuned_model = mmt.finetune(
    train_dataset,
    task_name="classification",
    epoch=10,
    lr=0.1
)
# Epoch 0: Train loss: 1.200
# Epoch 1: Train loss: 0.856
# ...
# Epoch 9: Train loss: 0.451

# Evaluate fine-tuned model
accuracy, embeddings, labels = mmt.evaluate(
    test_dataset, task_name="classification"
)
print(f"Test accuracy: {accuracy}")  # ~0.857

Real Example: ECG Classification

Complete workflow from moment_classification.ipynb:

from samay.model import MomentModel
from samay.dataset import MomentDataset
from samay.models.moment.momentfm.models.statistical_classifiers import fit_svm

# Initialize model
repo = "AutonLab/MOMENT-1-large"
config = {
    "task_name": "classification",
    "n_channels": 1,
    "num_class": 5
}
mmt = MomentModel(config=config, repo=repo)

# Load ECG5000 dataset
train_dataset = MomentDataset(
    name="ecg5000",
    path="data/ECG5000_TRAIN.csv",
    batchsize=64,
    mode="train",
    task_name="classification",
)

test_dataset = MomentDataset(
    name="ecg5000",
    path="data/ECG5000_TEST.csv",
    batchsize=64,
    mode="test",
    task_name="classification",
)

# Zero-shot: Extract embeddings and train SVM
train_accuracy, train_embeddings, train_labels = mmt.evaluate(
    train_dataset, task_name="classification"
)
test_accuracy, test_embeddings, test_labels = mmt.evaluate(
    test_dataset, task_name="classification"
)

clf = fit_svm(features=train_embeddings, y=train_labels)
train_acc = clf.score(train_embeddings, train_labels)
test_acc = clf.score(test_embeddings, test_labels)

print(f"Zero-shot Train accuracy: {train_acc:.2f}")  # 1.00
print(f"Zero-shot Test accuracy: {test_acc:.2f}")    # 0.93

# Fine-tune for better performance
finetuned_model = mmt.finetune(
    train_dataset,
    task_name="classification",
    epoch=10,
    lr=0.1
)

accuracy, embeddings, labels = mmt.evaluate(
    test_dataset, task_name="classification"
)
print(f"Fine-tuned Test accuracy: {accuracy}")  # 0.857

Visualizing Embeddings

Understand what the model learned with dimensionality reduction:

from sklearn.decomposition import PCA
import matplotlib.pyplot as plt

# Reduce embeddings to 2D
test_embeddings_manifold = PCA(n_components=2).fit_transform(test_embeddings)

plt.figure(figsize=(10, 7))
plt.title("ECG5000 Test Embeddings", fontsize=20)
plt.scatter(
    test_embeddings_manifold[:, 0],
    test_embeddings_manifold[:, 1],
    c=test_labels.squeeze(),
    cmap="viridis",
    alpha=0.7
)
plt.colorbar(label="Class")
plt.xlabel("Component 1")
plt.ylabel("Component 2")
plt.show()

Interpretation: Well-separated clusters indicate the model learned discriminative features for each class.

Advanced Techniques

Multi-Class Classification

For datasets with many classes:

config = {
    "task_name": "classification",
    "n_channels": 1,
    "num_class": 20  # e.g., gesture recognition with 20 gestures
}
mmt = MomentModel(config=config, repo=repo)

# Fine-tuning is highly recommended for many classes
finetuned_model = mmt.finetune(
    train_dataset,
    task_name="classification",
    epoch=15,  # More epochs for complex tasks
    lr=0.05
)

Multivariate Time-Series Classification

Classify sequences with multiple channels:

config = {
    "task_name": "classification",
    "n_channels": 6,  # e.g., accelerometer x, y, z + gyroscope x, y, z
    "num_class": 10   # 10 activity types
}
mmt = MomentModel(config=config, repo=repo)

train_dataset = MomentDataset(
    name="har",  # Human Activity Recognition
    path="data/HAR_TRAIN.csv",
    batchsize=32,
    mode="train",
    task_name="classification",
)

Imbalanced Classification

Handle class imbalance:

from sklearn.utils.class_weight import compute_class_weight
import numpy as np

# Compute class weights
class_weights = compute_class_weight(
    'balanced',
    classes=np.unique(train_labels),
    y=train_labels
)

# Use weighted SVM
from sklearn.svm import SVC

weighted_clf = SVC(class_weight='balanced', kernel='rbf')
weighted_clf.fit(train_embeddings, train_labels)

test_acc = weighted_clf.score(test_embeddings, test_labels)
print(f"Weighted SVM Test accuracy: {test_acc:.2f}")

Evaluation Metrics

Beyond Accuracy

For imbalanced datasets, use precision, recall, and F1:

from sklearn.metrics import classification_report, confusion_matrix
import seaborn as sns

# Predictions
y_pred = clf.predict(test_embeddings)

# Classification report
print(classification_report(test_labels, y_pred))

# Confusion matrix
cm = confusion_matrix(test_labels, y_pred)
plt.figure(figsize=(8, 6))
sns.heatmap(cm, annot=True, fmt='d', cmap='Blues')
plt.xlabel('Predicted')
plt.ylabel('Actual')
plt.title('Confusion Matrix')
plt.show()

Cross-Validation

Robust performance estimation:

from sklearn.model_selection import cross_val_score

# 5-fold cross-validation
scores = cross_val_score(
    clf, train_embeddings, train_labels, cv=5, scoring='accuracy'
)

print(f"Cross-validation scores: {scores}")
print(f"Mean accuracy: {scores.mean():.2f} (+/- {scores.std():.2f})")

Use Cases

Healthcare

ECG classification for arrhythmia detection, EEG seizure classification

Activity Recognition

Wearable sensor data for classifying human activities (walking, running, sitting)

Audio Classification

Speech command recognition, music genre classification

Industrial IoT

Equipment state classification (normal, faulty), predictive maintenance

Tips for Better Classification

Start with zero-shot SVM

MOMENT embeddings are powerful—often a simple SVM achieves 90%+ accuracy without fine-tuning.

Fine-tune for specialized domains

For medical, industrial, or rare event classification, fine-tuning improves accuracy by 5-15%.

Data augmentation

Use time-series augmentation (jitter, scaling, rotation) to improve generalization:

from tsaug import TimeWarp, Crop, Quantize

# Apply augmentation during training
augmented_data = TimeWarp().augment(train_data)

Ensemble methods

Combine predictions from multiple models (MOMENT + classical features) for robust classification.

Feature importance

Use attention weights or SHAP values to understand which time steps influence predictions.

Common Datasets

Dataset	Domain	Classes	Samples	Channels
ECG5000	Healthcare	5	5000	1
FordA	Automotive	2	4921	1
NATOPS	Gesture	6	360	24
UWaveGesture	Gesture	8	4478	3
SonyAIBORobot	Robotics	2	621	1

UCR/UEA Time Series Classification Archive provides 100+ benchmark datasets. Visit UCR Archive for more.

Next Steps

Need forecasting? See Zero-Shot Forecasting
Detect anomalies? Check Anomaly Detection Guide
Missing data in classification? Learn about Imputation

For more examples, explore the MOMENT Classification notebook.

Get Started

Core Concepts

Models

Guides

Time-Series Classification

Overview

Models Supporting Classification

Step-by-Step Workflow

Real Example: ECG Classification

Visualizing Embeddings

Advanced Techniques

Multi-Class Classification

Multivariate Time-Series Classification

Imbalanced Classification

Evaluation Metrics

Beyond Accuracy

Cross-Validation

Use Cases

Healthcare

Activity Recognition

Audio Classification

Industrial IoT

Tips for Better Classification

Common Datasets

Next Steps

Build docs developers (and LLMs) love

Get Started

Core Concepts

Models

Guides

​Overview

​Models Supporting Classification

​Step-by-Step Workflow

​Real Example: ECG Classification

​Visualizing Embeddings

​Advanced Techniques

​Multi-Class Classification

​Multivariate Time-Series Classification

​Imbalanced Classification

​Evaluation Metrics

​Beyond Accuracy

​Cross-Validation

​Use Cases

Healthcare

Activity Recognition

Audio Classification

Industrial IoT

​Tips for Better Classification

​Common Datasets

​Next Steps

Build docs developers (and LLMs) love

Overview

Models Supporting Classification

Step-by-Step Workflow

Real Example: ECG Classification

Visualizing Embeddings

Advanced Techniques

Multi-Class Classification

Multivariate Time-Series Classification

Imbalanced Classification

Evaluation Metrics

Beyond Accuracy

Cross-Validation

Use Cases

Tips for Better Classification

Common Datasets

Next Steps