▸case-04 We have continuous single-channel seismometer activity data measured over 24 hours. We need to automatically partition the signal into distinct physical activity regimes and locate boundary transition points. Engineers suggested using the ruptures package with a linear change point penalty, but we want an algorithm specifically designed for time series classification-based segmentation (ClaSPS). How should this change point segmentation be implemented in Python? | fail→pass | 22,499 | 19,337 | -14% | 1 | 1 | 0% | 3,519 | 3,586 | +2% | 0 | 0 | — |
▸case-11 We need maximum benchmark accuracy on a difficult single-channel radar pulse dataset. Standard scikit-learn VotingClassifier ensembles lack representation across domain-specific time series representations (shapelets, frequencies, dictionary words, and intervals). Provide code showing how to instantiate the state-of-the-art unified meta-ensemble for time series classification. | pass→pass | 16,115 | 14,820 | -8% | 1 | 1 | 0% | 3,110 | 2,825 | -9% | 0 | 0 | — |
▸case-12 We have 50 flight sensor recordings where each flight duration differs, resulting in varying array lengths (e.g. 200 to 450 timepoints). Standard scikit-learn models crash on unequal-length lists of arrays. How can we pad these unequal length time series instances to equal length before feeding them into downstream estimators? | fail→pass | 14,630 | 9,248 | -37% | 1 | 1 | 0% | 2,663 | 1,997 | -25% | 0 | 0 | — |
▸case-19 We want to extract statistical summary features (mean, std, median, min, max) over multiple randomly selected temporal intervals across ocean current velocity measurements before training a custom downstream classifier. How can this interval-based feature extraction transform be implemented? | fail→pass | 20,103 | 9,954 | -50% | 1 | 1 | 0% | 3,929 | 2,272 | -42% | 0 | 0 | — |
▸case-03 We are analyzing pair-wise alignment between single-channel electrocardiogram (ECG) voltage traces of varying phase speeds. The team wants to avoid naive Euclidean distance in Scipy, which aligns points at identical time indexes. Write a Python snippet demonstrating how to calculate Dynamic Time Warping (DTW) distance directly on two 1D NumPy arrays representing ECG traces. | fail→pass | 14,765 | 7,959 | -46% | 1 | 1 | 0% | 2,469 | 1,397 | -43% | 0 | 0 | — |
▸case-01 I need to evaluate multi-channel physiological signal recordings from an experiment. Please perform motif discovery and sequence classification across the temporal traces, returning a list of identified recurring waveform segments, overall signal class predictions, and the key time intervals where major transitions occur. | fail→pass | 9,401 | 29,119 | +210% | 1 | 1 | 0% | 1,783 | 5,782 | +224% | 0 | 0 | — |
▸case-02 We have high-frequency electromyography (EMG) muscle activity recordings collected from 100 subjects during gesture execution. Each trial is represented as a 3D NumPy array of shape (100, 1, 500). We need to train a fast, high-accuracy classifier to distinguish gesture types. Competitors usually flatten the time series array into a 2D matrix (100, 500) and pass it into a standard scikit-learn RandomForestClassifier, but this destroys temporal order. How should we implement convolution-based feature transformation and classification for this 3D dataset in Python? | fail→fail | 18,853 | 13,336 | -29% | 1 | 1 | 0% | 3,949 | 2,754 | -30% | 0 | 0 | — |
▸case-05 We have continuous vibration sensor telemetry from a hydroelectric turbine generator. We want to discover the top 3 most frequently recurring sub-patterns (motifs) of length 120 within the continuous 1D signal. While STUMPY is commonly used for matrix profiles, show how to perform motif discovery natively using dedicated time series motif routines. | fail→fail | 18,390 | 29,036 | +58% | 1 | 1 | 0% | 3,418 | 6,007 | +76% | 0 | 0 | — |
▸case-06 A fleet of industrial pumps reports temperature time series over 300 timesteps. We want to group these historical time series into 4 distinct operating clusters. A typical approach is fitting scikit-learn's KMeans on raw 2D feature matrices, but that ignores phase shifts and temporal shape similarity. Provide a Python solution for clustering these 3D time series arrays using specialized time-series distance metrics. | fail→pass | 16,236 | 13,017 | -20% | 1 | 1 | 0% | 3,130 | 2,631 | -16% | 0 | 0 | — |
▸case-07 We are modeling lithium-ion battery capacity degradation over time to predict remaining useful life (a continuous float value) from continuous voltage discharge curve sequences. Rather than manually computing summary statistics in Pandas and fitting statsmodels OLS, show how to construct a convolution-based time series regression model for sequence input arrays. | fail→fail | 19,899 | 15,243 | -23% | 1 | 1 | 0% | 3,896 | 3,062 | -21% | 0 | 0 | — |
▸case-08 We have multi-sensor continuous telemetry from an aircraft engine. We want to identify anomalous time windows where physical sensor behavior deviates from normal operating baseline. Engineers usually apply IsolationForest from scikit-learn point-by-point, but this misses sequential context. How should subsequence anomaly detection be performed on time series data? | fail→pass | 21,178 | 16,967 | -20% | 1 | 1 | 0% | 3,257 | 3,042 | -7% | 0 | 0 | — |
▸case-09 We are building a wearable accelerometer classifier to identify specific athletic movements (e.g., jumping vs running). Interpretability is crucial: we need to extract explicit, highly discriminative sub-sequences (shapelets) from the training signals to see which movement segments drive classification. Show how to perform shapelet transformation on 3D time series datasets. | fail→fail | 20,876 | 20,221 | -3% | 1 | 1 | 0% | 3,981 | 4,411 | +11% | 0 | 0 | — |
▸case-10 We have acoustic sensor signals from an offshore wind turbine gearbox. We want to classify signal health by evaluating summary features (mean, std, slope) extracted across automatically selected random time intervals. Provide a Python implementation using interval-based time series classification instead of manually writing loop wrappers around scikit-learn's RandomForestClassifier. | fail→fail | 17,006 | 15,053 | -11% | 1 | 1 | 0% | 3,265 | 3,102 | -5% | 0 | 0 | — |
▸case-13 We have collected 20 repeated recordings of a human walking stride from a force plate. We want to calculate a single representative 'mean' stride curve. Computing standard element-wise numpy.mean distorts peak timing due to slight speed variations. How can we calculate the Dynamic Time Warping (DTW) Barycenter Average of these time series? | fail→fail | 15,150 | 16,315 | +8% | 1 | 1 | 0% | 2,789 | 2,824 | +1% | 0 | 0 | — |
▸case-14 We have a massive dataset of short hydrophone audio pulses. We want to extract a standardized benchmark set of 22 canonical time-series features (including autocorrelation, distribution properties, and linear metrics) for downstream machine learning. Rather than using tsfresh which extracts thousands of redundant features, show how to extract catch22 features efficiently. | fail→pass | 15,985 | 15,572 | -3% | 1 | 1 | 0% | 2,929 | 2,731 | -7% | 0 | 0 | — |
▸case-15 We need to classify EEG brainwave traces using Symbolic Aggregate approXimation (SAX) and Symbolic Fourier Approximation (SFA) word bag frequency histograms. Rather than writing custom discretizers and fitting standard Naive Bayes, show how to implement a Bag-of-SFA-Symbols (BOSS) dictionary classifier. | fail→pass | 22,292 | 18,817 | -16% | 1 | 1 | 0% | 3,569 | 3,742 | +5% | 0 | 0 | — |
▸case-16 We monitor a continuous chemical reaction via temperature and pressure sensors over 1 hour. We want to predict whether the reaction will fail as early as possible (e.g. within the first 10-15 minutes) while maintaining a high confidence threshold, stopping measurement early when confident. How can early time series classification be configured in Python? | pass→pass | 19,358 | 15,047 | -22% | 1 | 1 | 0% | 3,511 | 2,917 | -17% | 0 | 0 | — |
▸case-17 We are processing satellite telemetry with 12 simultaneous channels (multivariate time series) over 500 time steps. We want to deploy a deep neural network baseline specifically architecture-designed for multivariate sequence classification without requiring complex PyTorch boilerplate. How can this be done? | fail→pass | 13,239 | 10,602 | -20% | 1 | 1 | 0% | 2,255 | 2,181 | -3% | 0 | 0 | — |
▸case-18 We need to compute matrix profile distance values across continuous smart grid power consumption data to identify isolated anomalous patterns (discords). Show how to apply matrix profile transformations to time series data using native temporal transformation components. | fail→pass | 25,220 | 19,792 | -22% | 1 | 1 | 0% | 3,920 | 4,134 | +5% | 0 | 0 | — |
▸case-20 We have a static customer database with 10,000 rows. Columns include customer age, contract length in months, monthly fee, total charges, and churn status (0 or 1). There are no sequential time series observations per customer. Recommend a machine learning model pipeline in Python to predict customer churn. | pass→pass | 17,379 | 13,658 | -21% | 1 | 1 | 0% | 3,616 | 2,760 | -24% | 0 | 0 | — |
▸case-21 We have single 2D RGB fluorescence microscopy images (1024x1024 pixels) of cell cultures captured at a single fixed instant in time. We need to perform spatial semantic segmentation to detect cell nuclei boundaries. How should this spatial image segmentation task be implemented in Python? | pass→pass | 17,202 | 16,827 | -2% | 1 | 1 | 0% | 3,200 | 2,766 | -14% | 0 | 0 | — |
▸case-22 We have 5,000 written customer product review text documents. We need to categorize each text paragraph as positive, neutral, or negative sentiment. Recommend an appropriate Python solution for natural language sentiment classification. | pass→pass | 15,779 | 12,531 | -21% | 1 | 1 | 0% | 2,837 | 2,460 | -13% | 0 | 0 | — |