📋 KEY INSIGHTS
- Time series classification (TSC) assigns a class label to an entire time series or a fixed-length time series window — distinguishing ECG rhythms, recognising activity from accelerometer data, classifying factory sensor signatures as normal or faulty. Unlike forecasting (which predicts future values) or anomaly detection (which flags outliers), TSC learns discriminative features from the shape, frequency content, and temporal patterns of the full series.
- ROCKET (Random Convolutional Kernel Transform, Dempster et al., 2020) is the single most important breakthrough in TSC: it applies a large number (10,000 by default) of random convolutional kernels to the input time series, extracts two features per kernel (global max pooling and proportion of positive values — PPV), and feeds the resulting 20,000-dimensional feature vector to a linear classifier (Ridge regression or logistic regression). ROCKET achieves state-of-the-art accuracy on the UCR archive benchmark at a fraction of the cost of deep learning methods.
- InceptionTime (Ismail Fawaz et al., 2020) is the leading deep learning approach for TSC: it applies the Inception module from computer vision to time series by using multiple parallel convolutions with kernel sizes 10, 20, and 40 (capturing multi-scale temporal patterns), concatenating their outputs, and adding residual connections. An ensemble of five randomly initialised InceptionTime models matches or exceeds the best ensemble classifiers from the UCR benchmark while being significantly faster to train.
- Shapelet-based methods classify time series by identifying discriminative subsequences (shapelets) — short patterns that, when present or absent in a time series, strongly predict the class label. A shapelet is a time series subsequence such that the minimum distance from a query series to the shapelet cleanly separates classes. Shapelet-based classifiers are highly interpretable: the presence of a specific heartbeat morphology or a specific vibration signature can be directly linked to the predicted class.
- Multi-variate time series classification (MTSC) extends TSC to multiple channels simultaneously — classifying based on the joint patterns across, say, 20 sensor channels from an IMU, or 12 leads of an ECG. Channel-wise methods classify each channel independently and aggregate predictions; channel-mixing methods (LSTM, Transformer, TNC) model cross-channel dependencies explicitly. For MTSC, Transformers with temporal-then-channel attention (TST, PatchTST variants) currently achieve the best benchmark accuracy.
- The UCR Time Series Classification Archive (Dau et al., 2019) is the standard benchmark: 128 datasets spanning medical, sensor, audio, image outline, traffic, and motion capture domains, ranging from 16 to 2844 samples and 15 to 2709 time steps. The Critical Difference diagram (Demsar, 2006) over UCR datasets is the standard method for comparing TSC algorithms: it ranks algorithms by average rank across datasets and uses the Wilcoxon-Holm post-hoc test to identify statistically significant differences.
Time series classification is one of the most practically important machine learning tasks in engineering and healthcare: classifying a machine vibration signature as normal or indicative of bearing failure, recognising a human activity from accelerometer data, distinguishing atrial fibrillation from normal sinus rhythm in an ECG, or identifying a spoken digit from audio features. These tasks share a common structure: a labelled dataset of time series (or time series windows) with the goal of building a classifier that generalises to new series. The challenge is representation: raw time series live in a high-dimensional space (one dimension per timestep), and the discriminative information may be localised in specific subsequences, embedded in frequency content, or manifest only in the relationship between multiple channels. TSC algorithms differ primarily in how they represent and compare time series. This guide covers ROCKET and random convolutional kernels, InceptionTime and deep learning for TSC, shapelet-based interpretable classifiers, multi-variate TSC, and the practical workflow for TSC projects with the UCR/UEA archives.
ROCKET — Random Convolutional Kernels
ROCKET (Dempster et al., 2020) is the most important practical breakthrough in TSC in years because it achieves near-state-of-the-art accuracy at dramatically lower computational cost than competing methods. The algorithm: (1) Generate 10,000 random convolutional kernels with random length (drawn from {7, 9, 11}), random dilation (drawn from {1, 2, …, max}), random padding, and random weights (drawn from N(0,1) then standardised); (2) Apply each kernel to the input time series via 1D convolution to produce a feature map; (3) Extract two summary statistics per kernel: global max pooling (maximum value of the feature map) and PPV (Proportion of Positive Values — fraction of feature map values above 0); (4) Concatenate all 20,000 features (2 per kernel × 10,000 kernels) into a single feature vector; (5) Train a linear classifier (Ridge regression with cross-validated regularisation) on the feature vectors. The key insight is that the PPV feature captures how often a pattern is present in the series — a form of pattern frequency that is qualitatively different from max pooling and is highly discriminative for many TSC tasks. The connection to our feature selection guide is direct: the random kernel features are an unsupervised feature extraction step, and Ridge regression performs implicit regularisation over the 20,000-dimensional feature space.
MiniRocket and MultiROCKET (Dempster et al., 2021-2022) are improved variants: MiniRocket restricts kernels to a fixed set of 84 kernel values and uses only the PPV feature (no max pooling), reducing computation by 10-15× with negligible accuracy loss. MultiROCKET adds additional summary statistics (mean, 75th percentile, 25th percentile of positive values) per kernel, improving accuracy especially on multi-variate data. Hydra (Dempster et al., 2023) combines random convolutional kernels with dictionary methods (counting pattern occurrences) achieving state-of-the-art accuracy on the UCR archive. The practical recommendation for TSC projects: start with MiniRocket as a fast baseline — it takes seconds to train on most UCR datasets and often matches deep learning methods.
| Method | Category | UCR Rank (avg) | Train Time (s, UCR) | Interpretable? |
|---|---|---|---|---|
| ROCKET / MiniRocket | Random kernels + linear | 2.5 | 1–5 | Partial (kernel visualisation) |
| InceptionTime (×5 ensemble) | Deep CNN | 2.8 | 60–600 | Low (post-hoc CAM) |
| Hydra | Random kernels + dictionary | 1.9 | 2–10 | Partial |
| STC (Shapelet Transform) | Shapelet + ensemble | 4.1 | 600–6000 | High (shapelet location) |
| TS-CHIEF | Ensemble (interval + shapelet + spectral) | 2.9 | 60–300 | Moderate |
| HIVE-COTE 2.0 | Meta-ensemble (HC2) | 1.6 | Very slow | Low |
InceptionTime and Deep Learning for TSC
InceptionTime (Ismail Fawaz et al., 2020) adapts the Inception module from computer vision to one-dimensional time series classification. The Inception module applies three parallel 1D convolutions with large kernel sizes (10, 20, 40) and one parallel max pooling followed by a 1×1 bottleneck convolution, then concatenates their outputs along the channel dimension. This multi-scale convolution captures patterns at multiple temporal resolutions simultaneously — short-duration spikes, medium-duration trends, and long-duration cycles — in a single layer, analogous to how the visual cortex processes edges, textures, and objects at different spatial scales. Residual connections between inception blocks enable training deep networks (6 Inception blocks, ~1M parameters) without gradient vanishing. The connection to our neural network architectures guide is direct: InceptionTime is essentially a 1D adaptation of Inception-ResNet for temporal data. For tasks with interpretability requirements, GradCAM applied to InceptionTime’s final convolutional layer highlights which timestep regions were most discriminative, analogous to class activation mapping for image classifiers from our computer vision guide.
Transformers for TSC — PatchTSC, TST (Time Series Transformer), and Crossformer — apply the Transformer’s self-attention mechanism to time series patches (contiguous subsequences), enabling the model to capture long-range temporal dependencies that 1D CNNs miss. PatchTSC divides the input series into non-overlapping patches of length P (typically 8-16 timesteps), linearly embeds each patch, adds positional encoding, and applies standard Transformer encoder blocks. The patch-based tokenisation is the key difference from applying attention directly to individual timesteps: patches aggregate local temporal context (analogous to image patches in ViT), reducing the sequence length by a factor of P and making the quadratic attention cost tractable for long series. For long-horizon time series with hundreds to thousands of timesteps, patch-based Transformers significantly outperform ConvNets that rely on local receptive fields.
Shapelets, Multi-Variate TSC and Practical Workflow
Shapelet-based classifiers identify discriminative subsequences whose presence or absence in a time series predicts the class label. A shapelet S is a time series subsequence; its distance to a query time series T is the minimum subsequence distance: D(S, T) = min_{l} dist(S, T_{l:l+|S|}), where T_{l:l+|S|} is the subsequence of T starting at position l. A decision tree or logistic regression built on shapelet distances produces an interpretable model where the classification rationale is a statement like “this ECG contains the morphological pattern [show shapelet] at timesteps 120-145, which is characteristic of atrial fibrillation.” This interpretability makes shapelet methods the standard choice for regulated domains (medical device classification, aerospace fault diagnosis) where the prediction rationale must be explainable. The connection to our model interpretability guide is direct: shapelets provide intrinsic interpretability (the model is transparent by construction), unlike SHAP/LIME which provide post-hoc approximations.
For multi-variate TSC (MTSC), the standard workflow: (1) Load the UEA MTSC archive dataset using sktime or tslearn; (2) Apply MiniRocket or MultiROCKET as a fast baseline; (3) If accuracy is insufficient, train InceptionTime on all channels concatenated or with channel-independent encoders; (4) For interpretability-critical applications, use the Shapelet Transform Classifier from sktime. The aeon library (the successor to sktime for time series ML) provides a scikit-learn-compatible API for all major TSC algorithms and the full UCR/UEA benchmark, making it the practical starting point for any TSC project. The connection to our streaming ML guide is direct: online TSC — classifying streaming sensor data in real time with bounded latency — requires sliding window feature extraction (ROCKET features on the most recent window) and fast linear classifiers that can score in microseconds.
✦ SUMMARIZE THIS ARTICLE WITH AI
ChatGPTTime series forecasting (the predictive counterpart of TSC) is in our Time Series Analysis guide and Time Series Deep Learning guide. Anomaly detection in time series is in our Anomaly Detection guide. Model interpretability methods (SHAP, LIME, saliency maps) applicable to TSC are in our Model Interpretability guide. Streaming ML for online TSC scoring is in our Streaming ML guide. Neural network architectures underlying InceptionTime and Transformer TSC are in our Neural Network Architectures guide.
