Unlocking Time Series Insights with TSFresh: A Python Guide

Time series analysis is a powerful tool in data science, allowing us to understand the underlying patterns in temporal data and make predictions. One of the challenges in working with time series data is extracting meaningful features that can be used for machine learning. This is where tsfresh comes into play.

Introduction to TSFresh

Time series analysis is crucial in various domains like finance, healthcare, and retail. Traditional methods often involve manual feature extraction, which is not only time-consuming but also prone to human error and bias. Enter TSFresh (Time Series Feature extraction based on scalable hypothesis tests), a Python library that automatically extracts hundreds of features from time series data, offering a more efficient and objective approach.

TSFresh stands out for its ability to handle time series datasets of varying lengths and frequencies. It automatically identifies and extracts relevant characteristics from the data, such as trends, seasonality, and autocorrelation. This level of automation and detail in feature extraction was not as readily available in previous methods.

1. Formatting Data for TSFresh

Before diving into TSFresh, it's essential to format your data correctly. TSFresh requires a specific structure where each row represents an observation and each column a time step. Here's how to prepare your data:

Example Data Preparation

Suppose you have a wide DataFrame df from a CSV file:

import pandas as pd

# Load the CSV file
df = pd.read_csv('your_timeseries_data.csv')

# Display the first few rows
print(df.head())
        

This data needs to be transformed into a long format where one column contains all the time series identifiers, another the time stamps, and the last one the observed values.

# Transforming into long format
long_df = df.melt(id_vars=['Time_Series_ID'], var_name='Time', value_name='Value')

# Display the transformed data
print(long_df.head())
        

You may also use stack()

import pandas as pd


# Assuming `df` is your original DataFrame
df = pd.DataFrame({
    # Your data here, with 'Time' as one of the columns
})

# Set 'Time' as the index
df.set_index('Time', inplace=True)

# Convert the DataFrame from wide to long format
df_long = df.stack().reset_index()

# Rename the columns to match tsfresh format
df_long.rename(columns={'level_1': 'id', 0: 'value'}, inplace=True)        

At the end, you will need the following format:

2. Extracting Features with TSFresh

After preparing your data, the next step is to extract features using TSFresh.

Feature Extraction

from tsfresh import extract_features

# Extract features
extracted_features = extract_features(long_df, column_id='Time_Series_ID', column_sort='Time')

# Display extracted features
print(extracted_features.head())
        

Be prepared to see huge number of features. I usually get in hundreds. Here is the shape from one of my projects with 783 features

3. Understanding Extracted Features

TSFresh extracts a wide array of features. These include basic statistics like mean and median, as well as more complex ones like Fourier transforms and autocorrelation. Understanding these features involves recognizing the type of information each feature represents about the time series.

Exploring Features

You can explore the features using descriptive statistics and visualizations:

# Descriptive statistics
print(extracted_features.describe())

        

# Visualization (for example, using seaborn)
import seaborn as sns
sns.pairplot(extracted_features)
        

4. Reducing Features to the Most Important Ones

Not all extracted features are equally important. TSFresh allows for feature selection, reducing the feature set to those

most relevant for your specific problem.

Feature Selection

TSFresh offers methods for filtering out irrelevant features based on their importance scores. This can be done using the select_features function, which considers the relevance of each feature to the target variable.

Here's an example of how to use it:

from tsfresh import select_features
from tsfresh.utilities.dataframe_functions import impute

# Impute missing values
impute(extracted_features)

# Assume 'y' is your target variable
y = [1, 0, 1, 0, 1]  # Example binary target

# Selecting important features
important_features = select_features(extracted_features, y)

# Display important features
print(important_features.head())
        

In this example, y represents the target variable you are trying to predict or classify. The select_features function filters out the irrelevant features, keeping only those with significant predictive power.

I was able to reduce features from 800 to 500

1. Time Series Features: Many of these features are derived from time series analysis, which is crucial in monitoring activity. Features like value__maximum, value__absolute_maximum, and value__mean_change indicate statistical properties of the sensor readings over time. For example, value__maximum might indicate the maximum sensor reading in a given period, which could be crucial for identifying abnormal activity.
2. Change Quantiles: Features like value__change_quantiles__... provide information about the distribution of changes in sensor readings. In the context of volcanic activity, sudden or significant changes might be indicative of upcoming eruptions.
3. Time Reversal Asymmetry Statistic: Features like value__time_reversal_asymmetry_statistic__lag_1 could be helpful in identifying non-linear and complex patterns in the sensor data, which are common in geophysical time series.
4. Linear Trends: Features such as value__agg_linear_trend__... detect trends in the data over time. A changing trend might indicate a buildup towards an eruption.
5. Fourier Transform Coefficients: The value__fft_coefficient__... features represent the frequency domain of the sensor signals. Changes in frequency patterns can be indicative of shifts in volcanic activity.
6. Wavelet Features and Peaks: Features like value__number_cwt_peaks__n_1 and value__number_peaks__n_50 measure the number of

peaks in the wavelet transform and the raw signal, respectively. These can be important for identifying sudden spikes or anomalies in sensor readings, which might be signs of increasing volcanic activity.

1. Autoregressive Coefficients: value__ar_coefficient__coeff_0__k_10 relates to the autoregressive model of the time series. This could indicate how past values of the sensor readings are influencing current readings, which is important in understanding the progression of volcanic activity.
2. Linear Trend Intercept: The value__linear_trend__attr_"intercept" might provide insight into the baseline level of the sensor readings, which could be crucial for establishing normal versus abnormal activity levels.

Conclusion

TSFresh is a powerful tool for automatic feature extraction in time series analysis. By automating the extraction process, it saves time and reduces the potential for human error, allowing analysts to focus more on modeling and interpreting results. This guide provided a simple and detailed walkthrough of formatting data for TSFresh, extracting features, understanding them, and finally reducing them to the most relevant ones. Through practical examples and clear explanations, we hope to have unlocked the potential of TSFresh for your time series analysis projects.

Remember, time series analysis is a complex field, and TSFresh is just one of the tools at your disposal. Combining its capabilities with your domain knowledge and other data science techniques can lead to more insightful and accurate analyses. Happy analyzing!