---
title: 'MiniROCKET: Deterministic Time Series Classifier'
url: https://www.emergentmind.com/topics/minirocket-0b85e58c-506c-414c-bb36-f5cdef49c5d3
type: topic
---

# MiniROCKET: Deterministic Time Series Classifier

MiniROCKET is a transform for time series classification that converts each input time series into a high-dimensional feature vector using many simple convolutions, then trains a linear classifier on those features. It is a reformulation of ROCKET designed to preserve ROCKET’s key strengths—especially dilation and PPV pooling—while removing almost all randomness and dramatically reducing runtime. In the original evaluation, it is up to \(75\) times faster on larger datasets, often marginally more accurate than ROCKET, and makes it possible to train and test a classifier on all \(109\) datasets from the UCR archive to state-of-the-art accuracy in less than \(10\) minutes [2012.08791].

## 1. Origins and position within the ROCKET family

MiniROCKET emerged as a speed-oriented reformulation of ROCKET. ROCKET uses \(10{,}000\) random kernels, computes two features per kernel—global max and PPV—and produces \(20{,}000\) features per series. Its kernels are random in length, weights, bias, dilation, and padding. MiniROCKET keeps the basic transform-plus-linear-classifier architecture, but replaces most of this stochasticity with a much smaller, mostly fixed design [2012.08791].

The original paper identifies dilation and PPV as the two most important aspects of ROCKET, and MiniROCKET is organized around that observation. Rather than preserving ROCKET’s full random kernel space, it fixes the kernel family, fixes the dilation schedule relative to input length, fixes padding alternation, and derives biases from actual convolution outputs rather than from a fixed uniform distribution. This makes the method “almost deterministic,” with the only stochastic part in the default version being the selection of the training example used to generate bias quantiles [2012.08791].

In later literature, MiniROCKET is treated both as a strong baseline and as a modular component. Some works keep the original transform and change only pooling or feature selection; others use MiniROCKET as a fixed morphology encoder, a pruning target, or a deployable edge classifier. This suggests that MiniROCKET’s historical importance lies not only in benchmark performance, but also in establishing a practical design pattern: deterministic or nearly deterministic random-kernel feature extraction followed by shallow classification [2202.08055].

## 2. Transform design and feature construction

MiniROCKET uses a very small, fixed family of convolutional kernels. All kernels have length \(9\). Each kernel uses only two values,
\[
\alpha = -1, \qquad \beta = 2,
\]
and MiniROCKET uses only kernels with exactly three \(\beta\) values and six \(\alpha\) values. The number of distinct kernels is therefore
\[
\binom{9}{3} = 84.
\]
Because each kernel has six \(-1\)s and three \(2\)s, the weights sum to zero,
\[
6(-1) + 3(2) = 0,
\]
and, more generally, \(\beta = -2\alpha\). The zero-sum property implies invariance to adding or subtracting a constant \(c\) from the input, so the kernels respond to relative magnitudes rather than absolute level shifts [2012.08791].

For an input time series
\[
X = [x_0, x_1, \ldots, x_{n-1}],
\]
kernel
\[
W = [w_0, w_1, \ldots, w_{m-1}],
\]
and dilation \(d\), the convolution is written as
\[
X * W_d = \sum_{j=0}^{m - 1} x_{i - (\lfloor \tfrac{m}{2} \rfloor \cdot d) + (j \cdot d)} \cdot w_{j}, \quad \forall i \in \{0,1,\ldots,n-1\}.
\]
MiniROCKET retains only PPV, not global max. If \(C = X * W - b\), then
\[
\mathrm{PPV}(C) = \frac{1}{n} \sum [c > 0].
\]
Equivalently,
\[
\mathrm{PPV}(X * W - b) = \frac{1}{n} \sum [X * W - b > 0].
\]
The paper explicitly interprets PPV as an empirical CDF evaluation of the convolution output at threshold \(b\) [2012.08791].

Biases are one of the defining differences from ROCKET. ROCKET samples bias from \(\mathcal{U}(-1,1)\), whereas MiniROCKET samples quantiles from actual convolution outputs. Multiple PPV features are then generated from one convolution output by evaluating several bias thresholds. The default target is about \(10{,}000\) features; because the \(84\)-kernel structure is fixed, the practical default is \(9{,}996\) features. The default maximum number of dilations per kernel is \(32\), and features are allocated so that smaller dilations receive more features, matching ROCKET’s effective bias toward shorter scales [2012.08791].

The downstream classifier remains linear. In the original paper, smaller datasets use a ridge regression classifier from scikit-learn, while larger datasets use logistic regression trained with Adam in PyTorch. MiniROCKET therefore remains a transform-based classifier rather than a learned end-to-end convolutional network [2012.08791].

## 3. Computational properties, determinism, and scaling

MiniROCKET’s main practical contribution is runtime reduction without a corresponding loss of accuracy. Its asymptotic complexity remains in the same class as ROCKET,
\[
O(k \cdot n \cdot l_{\text{input}}),
\]
but constant factors are much smaller. On the \(109\) UCR datasets, single-core total compute time drops from \(2\)h\(02\)m for ROCKET to \(8\)m for MiniROCKET, and transform time drops from \(1\)h\(55\)m to \(2\)m\(30\)s. Average total transform time is more than \(30\times\) lower, and on large datasets the speedup reaches \(75\times\) [2012.08791].

Several design choices enable this reduction. First, the kernel family is fixed at \(84\) kernels rather than sampled anew. Second, MiniROCKET computes only PPV, immediately halving feature extraction work relative to ROCKET’s PPV-plus-max formulation. Third, it reuses one convolution output to generate multiple features via multiple biases. Fourth, the restricted \(\{-1,2\}\) weight set allows the transform to be implemented largely through additions and structured reuse. The original implementation exploits the decomposition \(2=-1+3\), precomputes \(A=-X\) and \(G=3X\), and shares a large fraction of the work across all \(84\) kernels for a given dilation [2012.08791].

The method is “almost deterministic” because only one default step remains stochastic: selecting the training example used to generate bias quantiles for a kernel/dilation combination. The paper also defines a fully deterministic variant that uses convolution outputs over the entire training set rather than a single sampled example, but this incurs noticeably higher computational and memory cost and yields negligible accuracy differences in practice [2012.08791].

The original paper also emphasizes that MiniROCKET does not require input normalization in the usual sense. The zero-sum kernels and bias thresholds sampled from actual convolution outputs make feature construction scale-aware. A practical caveat, however, is that the original paper focuses on univariate TSC and mentions only a naive facility for multivariate data in the repository, leaving later multivariate use to external implementations or domain-specific adaptations [2012.08791].

## 4. Variants, extensions, and pruning methods

Later work extends MiniROCKET along several distinct axes. One line revisits the pooling stage. HDC-MiniROCKET reformulates MiniROCKET’s PPV pooling as a bundling operation in hyperdimensional computing, then adds explicit timestamp binding. In that formulation, MiniROCKET is the special case \(s=0\). On a synthetic dataset in which the class depends on whether a sharp peak occurs in the first or second half of the sequence, MiniROCKET achieves \(65.0\%\) accuracy whereas HDC-MiniROCKET with \(s=1\) reaches \(97.0\%\); on a harder subset, the gap is \(56.8\%\) versus \(94.1\%\). On the \(128\) UCR datasets, an oracle over \(s \in \{0,\dots,6\}\) improves over MiniROCKET on \(81\) datasets, with average improvement \(3.1\%\) across those improved datasets [2202.08055].

A second line varies the feature semantics while retaining MiniROCKET’s kernel machinery. SelF-Rocket treats MiniROCKET as the special case
\[
f(X,\{I\},PPV),
\]
then searches over input representations \(\{I, DIFF\}\) and pooling operators \(\{PPV, GMP, MPV, MIPV, LSPV\}\). On \(112\) UCR datasets and \(30\) resamples, the reported mean accuracy gain over MiniROCKET is about \(1.23\%\) [2409.01115].

A third line prunes MiniROCKET after feature extraction. Detach-ROCKET applies sequential feature detachment to MiniROCKET features; in the reported MiniROCKET experiment, retaining \(10\%\) of the original features yields an average relative accuracy change of about \(-0.38\%\) [2309.14518]. POCKET instead formulates pruning as a group elastic net classification problem, where one group corresponds to one kernel in MiniROCKET. On the first \(30\) UCR datasets, the original MiniROCKET average accuracy is \(84.37\%\); POCKET Stage 1 reduces this to \(83.48\%\), and Stage 2 recovers it to \(84.12\%\), while pruning about \(64\%\) of kernels on average [2309.08499].

Other work changes the input representation rather than the kernel transform itself. ROMAN treats MiniROCKET as a pooled random-convolution classifier and rewrites coarse temporal position and scale into pseudochannels before the transform. On synthetic tasks, MiniROCKET improves from \(0.710\) to \(0.908\) on coarse position awareness, from \(0.514\) to \(0.676\) on long-range correlation, and from \(0.547\) to \(0.735\) on multiscale interaction, but falls from \(0.966\) to \(0.653\) on full positional invariance [2604.02577].

Finally, HIT-ROCKET replaces MiniROCKET’s kernel family with Hadamard-vector kernels. It is explicitly presented as “an improvement and supplement to ROCKET-family methods, especially miniROCKET,” and claims \(50\%\) shorter training time than miniROCKET under identical hyperparameters, higher accuracy once the feature dimension exceeds \(1\)K, and smaller F1 variance under additive noise [2511.01572].

## 5. Representative applications across domains

MiniROCKET has been applied far beyond the univariate UCR setting. In interactive performance, it serves as the time-series recognition engine in a personalized dance system built from four wireless IMUs on both wrists and both ankles. The system streams \(24\) channels at \(48\) Hz, segments motion into \(2\)-second windows, applies MiniROCKET followed by a ridge-based linear classifier, and reports \(96.05\%\) mean accuracy, \(96.62\%\) macro-averaged F1, all per-class AUC scores above \(0.99\), about \(15\) ms inference time, and under \(50\) ms end-to-end latency [2511.02351].

In wearable human activity recognition, MiniROCKET is evaluated on the UCI smartphone HAR dataset. On the precomputed \(561\)-feature representation, it achieves \(0.9881 \pm 0.0031\) accuracy, \(0.9886 \pm 0.0029\) F1, and \(0.9932 \pm 0.0018\) AUC. On raw one-channel sensor sequences without preprocessing, the best result is on total acceleration \(y\)-axis, with \(0.9350 \pm 0.0054\) accuracy, \(0.9388 \pm 0.0051\) F1, and \(0.9633 \pm 0.0031\) AUC [2402.18296].

Clinical and biomedical uses are similarly diverse. In automated identification of Action Research Arm Test items from wrist-worn IMUs, MiniROCKET reaches \(82.3\%\) accuracy for ARAT domain classification on the full left-wrist dataset and \(47.2\%\) for \(20\)-class item recognition; after removing the longest \(25\%\) of sequences, domain classification rises to \(93.9\%\) [2504.12921]. In ECG-RAMBA, MiniROCKET is the fixed morphology branch: each \(5\) s, \(12\)-lead ECG slice yields \(10{,}000\) deterministic features, which are compressed by fold-aware PCA to \(3072\) dimensions and fused with HRV and Mamba-based context. The full system achieves macro ROC-AUC \(\approx 0.85\) on Chapman--Shaoxing and PR-AUC \(=0.708\) for zero-shot atrial fibrillation detection on CPSC-2021, compared with \(0.382\) for a comparable raw-signal Mamba baseline [2512.23347]. In four-class motor imagery EEG classification on PhysioNet, MiniROCKET reaches \(98.63\%\) mean accuracy, slightly above a CNN-LSTM baseline at \(98.06\%\) [2508.16179].

Industrial and edge uses show a different aspect of the method. For ATM predictive maintenance from event logs, MiniROCKET transforms daily \(38\)-dimensional multivariate time series of length \(144\) and, with a Ridge classifier, yields accuracy \(0.7286_{0.042}\), balanced accuracy \(0.6639_{0.024}\), F1 \(0.3113_{0.046}\), AUC \(0.6639_{0.024}\), and time \(23.3_{7.3}\) s; it significantly outperforms InceptionTime on the imbalance-aware metrics reported [2305.10059]. In a TinyML tool-usage classifier running directly on a retrofit BLE sensor tag, MiniROCKET with \(84\) kernels and \(84\) features achieves \(96.9\%\) accuracy / F1 \(=0.969\), occupies \(7\) kB of flash and \(3\) kB of RAM, runs in \(8.6\) ms, and supports an average system power below \(15\,\mu\text{W}\) [2310.14758].

MiniROCKET also appears in unsupervised scientific workflows and foundation-model pipelines. In astronomical orbital clustering, it transforms \(22{,}288\) normalized \(400\)-step resonant-angle series into a \(9{,}996\)-dimensional feature space; MiniROCKET-only clustering gives silhouette \(0.5656\), DB \(0.6593\), CH \(42851.35\) for \(\varphi_1\), and silhouette \(0.6228\), DB \(0.5108\), CH \(89090.91\) for \(\varphi_2\), while the top-performing pipelines all include MiniROCKET [2603.13177]. In cross-species antimicrobial resistance prediction from genomic foundation-model embeddings, MiniROCKET summarizes ordered \(41\)-channel token streams with \(2000\) kernels over chunks of length \(2048\); on the ampicillin val\(_\text{outside}\) split, MiniRocket with cosine \(k\)-NN reaches AUROC \(0.926\), AUPRC \(0.992\), F1 \(0.982\), and MCC \(0.753\), whereas a comparable global-pooling \(k\)-NN model reaches AUROC \(0.515\), AUPRC \(0.902\), F1 \(0.901\), and MCC \(0.148\) [2603.11141]. In hyperspectral spectral classification, MiniROCKET is reported to dominate below about \(100{,}000\) training samples per class and, on HYPSO-1 with the full dataset, achieves OA \(74.79\), AA \(78.15\), F \(72.27\), and mIoU \(58.91\), above the corresponding 1D-Justo-LiuNet values \(72.90\), \(76.51\), \(69.69\), and \(56.37\) [2509.13809].

## 6. Limitations, failure modes, and recurrent misunderstandings

A common misunderstanding is to treat MiniROCKET as a sequence model. In several later systems it is explicitly not the temporal reasoning module. ECG-RAMBA is a clear example: MiniROCKET supplies a frozen, deterministic per-window morphology representation, while long-range context is handled by a bi-directional Mamba backbone. The paper is explicit that Mamba does not operate on a sequence of MiniRocket features [2512.23347]. This suggests that MiniROCKET’s core abstraction is a fast local-pattern transform, not an end-to-end contextual model.

A second limitation is its treatment of global temporal position. HDC-MiniROCKET was motivated precisely by the claim that MiniROCKET can “systematically overcome catastrophic failures” only when augmented with explicit time encoding. The synthetic early-peak versus late-peak task is the canonical example: MiniROCKET can detect the local event, but PPV pooling may wash out where it occurred. ROMAN reaches a similar conclusion from the representation side: MiniROCKET benefits when coarse position, long-range relations, or multiscale interactions matter, but suffers when full positional invariance is the correct inductive bias [2202.08055; 2604.02577].

A third limitation is that MiniROCKET is not universally the best surrounding structure for every problem. In hierarchical time-series classification, the reported advantage is narrow: hierarchical classification is significantly better than flat classification only when MINIROCKET is paired with Task Similarity Distance, whereas STSF and SVM continue to favor flat classification across the studied settings [2402.05275]. In other words, MiniROCKET’s behavior can depend strongly on how class structure is represented around it.

Finally, many applied papers leave parts of the MiniROCKET pipeline under-specified. The dance-recognition study does not report MiniROCKET hyperparameters, overlap or stride, normalization, or exact augmentation settings; the smartphone HAR study does not document kernel counts, downstream classifier details, or how the \(561\)-feature vectors are arranged for MiniROCKET input; the ARAT study gives the PPV formulation and multivariate `sktime` implementation, but omits the exact downstream classifier [2511.02351; 2402.18296; 2504.12921]. A plausible implication is that MiniROCKET’s practical reputation partly rests on its robustness to incomplete tuning information: even when full implementation detail is absent, the transform often remains effective. That implication, however, should be separated from full reproducibility, which these papers do not provide.

Source: https://www.emergentmind.com/topics/minirocket-0b85e58c-506c-414c-bb36-f5cdef49c5d3