Time series classification

Overview

Time series classification assigns one label to a whole sequence, such as a sensor trace or a GPS log, by looking at how its values change over time 2102.00457 §1 1804.02386 §1. It differs from time series forecasting, which predicts future values, and from dynamic time warping, a distance between two sequences that can also be used to label one by its nearest reference.

The fourteen cards behind this page cover twelve methods or applications, a tutorial and a benchmark:

• A 1-D convolutional network over four hand-computed motion channels per GPS point 1804.02386 §3.1 1804.02386 §3.2.
• A fixed, non-learned feature recipe: random convolution kernels, pooled with four operators, followed by a linear classifier 2102.00457 §3.
• A meta ensemble of four classifiers, each on its own representation, whose class probabilities are combined with weights from training accuracy 2104.07551 §3 2104.07551 §3.1 2104.07551 §3.4.
• A Transformer encoder on raw per-date satellite reflectance, pooled over time into one vector for crop-type classification 1910.10536 §4.2 1910.10536 §7.
• Self-supervised pre-training of a 1-D CNN on ECG, where the positives are segments from the same patient, and the frozen or fine-tuned encoder is then used for arrhythmia classification 2005.13249 §4.2 2005.13249 §5.3.
• Temporal and contextual contrasting of two augmented views, followed by a linear probe on HAR, Sleep-EDF and Epilepsy data. Positives are views of the same sample, so no patient labels are needed 2106.14112 §3.
• A Transformer over (time, variable, value) triplets for sparse, irregularly sampled clinical records, pretrained by forecasting and fine-tuned for in-hospital mortality 2107.14293 §3.1 2107.14293 §3.2.6.
• Deterministic counts of short symbolic words taken from many sliding windows, fed to a linear classifier. Each count can be traced back to the signal 2311.18029 §4.2 2311.18029 §5.
• Gramian angular fields and Markov transition fields that turn each series into an image, classified by a Tiled CNN 1506.00327 §2 1506.00327 §3.3.
• The Transformer encoder, which the tutorial explains and adapts to series data. It has a classification head in one model it surveys 2205.01138 §2 2205.01138 §4.3.
• Hand-computed statistics per channel, fed to an SVM, or raw padded series fed to an LSTM, for classifying which object a robot squeezed. Recordings come from one device at a time 2204.06343 §4.1 2204.06343 §4.2.
• A 1-D CNN on 24 channels of a robot arm's joint and force readings, which labels one of six glovebox tasks 2402.10220 §IV-A 2402.10220 §IV-B.
• A benchmark rather than a new method: the best algorithm from each of eight families, run on the UCR archive under one protocol 2304.13029 §3 2304.13029 §4.
• A Random Forest on the four spectral bands of satellite cells, one model per year, labelling each cell as totally burned, partially burned or non-burned. The paper is mainly a dataset for one national park, with this classifier as its baseline 2609.23687 §1 2609.23687 §2.2.15.

The evidence is uneven. Each method card reports author results on its own tasks and protocol. The tutorial runs no experiments and gives no numbers for its classification examples 2205.01138 §4.5. The benchmark gives the most directly comparable numbers, but only for univariate series 2304.13029 §3. The two robot cards are small studies with author-reported accuracy only. One has about 24,000 squeezes across four grippers and hands; the other has 24 to 200 recordings per dataset 2204.06343 §1 2402.10220 §III-D. The HC2 card is the largest study on UCR and UEA problems, but its runs take hundreds of hours. The satellite card is one crop-mapping study in Bavaria, with an industry-preprocessed data set whose loader the card could not reproduce from the released repository 2104.07551 §4 1910.10536 §7.2.2 (see the card). The burned-area card compares one run per year with a satellite burned-area product for the same park, and that comparison rests on a single run 2609.23687 tab. 11. Its labels were drawn by hand for one park over three years 2609.23687 §2.2.13 2609.23687 §3.3.

Variants

1-D CNN on engineered per-point channels

Each 200-point segment becomes a 1 × 200 × 4 input. The four channels are speed, acceleration, jerk and bearing rate, all computed from the GPS points 1804.02386 §3.2.1 1804.02386 eq. 1–7. The network stacks 1 × 3 convolutions with ReLU, 1 × 2 max-pooling, and a softmax over five travel modes 1804.02386 §3.2 1804.02386 §3.2.5. The paper reports 79.8% test accuracy for one model, and 84.8% for an ensemble of seven copies trained on bootstrap samples 1804.02386 tab. 4. The card for this variant is Inferring transportation modes from GPS trajectories using a convolutional neural network.

The inputs and their preparation are part of the method. The paper removes points that exceed per-mode speed or acceleration limits and smooths the track with a Savitzky–Golay filter 1804.02386 §3.1 1804.02386 tab. 2. A hand-tuned baseline of classical classifiers on the same segment features is reported in 1804.02386 tab. 6: a random forest reaches 78.1% and a decision tree 75.2%.

Random convolution kernels with a linear classifier (MultiRocket)

MultiRocket keeps the fixed-kernel design of MiniRocket and adds two changes. It convolves the raw series and its first-order difference, so there are two views 2102.00457 §3.1. It pools each convolution output with four operators: the proportion of positive values, the mean of the positive values, their mean index, and the longest run of positive values 2102.00457 §3.4. The default gives 49,728 features, and a ridge classifier is fitted on them 2102.00457 §3.5. The 84 base kernels are fixed, and only the dilations and biases are fitted from training data 2102.00457 §3.2. The card for this variant is MultiRocket.

Across 109 univariate UCR datasets with 30 resamples, the authors report that MultiRocket is significantly more accurate than MiniRocket, and not significantly less accurate than HIVE-COTE 2.0 2102.00457 §4.1. The paper covers univariate series only 2102.00457 §5. See random convolutional kernels for the shared idea.

Meta ensemble of four specialists (HIVE-COTE 2.0)

HIVE-COTE 2.0 (HC2) trains four modules on the same training set: the Shapelet Transform Classifier, the Temporal Dictionary Ensemble, the Diverse Representation Canonical Interval Forest and Arsenal, an ensemble of ROCKET-style classifiers 2104.07551 §3 2104.07551 §3.1 2104.07551 §3.2 2104.07551 §3.3 2104.07551 §3.4. Each module also gives an estimate of its own accuracy from the training data. The module probabilities are combined with weights that are a power of that estimate, with exponent 4 2104.07551 §3. The paper describes this weighting in words and gives no formula, so the combination rule is the authors' description and not a published equation 2104.07551 §3. The card for this variant is HIVE-COTE 2.0: a new meta ensemble for time series classification.

On 112 univariate UCR problems, with 30 resamples each, the authors report that HC2 is significantly more accurate than HC1, InceptionTime, ROCKET and TS-CHIEF 2104.07551 §5.1. Its mean accuracy difference against ROCKET is −2.49%, with HC2 winning on 97 problems and losing on 11 2104.07551 tab. 3. On 26 multivariate UEA problems, the paper reports significant gains over DTW-D, ROCKET, HC1 and CIF 2104.07551 §5.2. The paper's ablations show the full combination significantly beats each of 11 variants built from its four components 2104.07551 §6 2104.07551 tab. 7.

The cost is the main constraint. One sequential HC2 run over one resample of the 112 UCR problems takes about 340 hours, against about 2.9 hours for ROCKET 2104.07551 tab. 4. HC2 can run under a time contract. On the five problems where a full build takes over 12 hours, a 4-hour contract reaches 98% of full-build accuracy and a 12-hour contract reaches 99% 2104.07551 §7 2104.07551 tab. 8. The paper does not say whether those contract results include the time spent on accuracy estimates, and the card's Paper vs code section records that the released code does not count them against the contract 2104.07551 §7. The reference toolkit is GPL-3.0, so cite it rather than copy its code.

Contrastive pre-training of a 1-D CNN, then classification (CLOCS)

CLOCS trains a 1-D CNN without labels, then reuses it for a labelled task. Its inputs are 2,500-sample frames of ECG, and the encoder is three convolution blocks followed by a linear layer to an embedding of size E 2005.13249 §5.5 2005.13249 app. C.1. Unlike a per-instance contrastive method, two segments count as a positive pair when they come from the same patient 2005.13249 §4.1. There are three variants: adjacent non-overlapping segments (CMSC), different leads (CMLC), or both (CMSMLC) 2005.13249 §4.2. The loss is an InfoNCE-style cosine objective at temperature 0.1, summed over both directions 2005.13249 eq. 1–4 2005.13249 §4.3. The downstream task is cardiac arrhythmia classification, scored by test AUC, with either a frozen encoder (linear evaluation) or a fine-tuned one 2005.13249 §5.3.

The paper reports that CMSC gives the best linear AUC on Chapman at E=128, F=0.5 and 4 leads: 0.896 against 0.738 for SimCLR, with five seeds 2005.13249 tab. 1. The gain is 0.158 in AUC points, which the authors call 15.8% 2005.13249 §6.2. On PhysioNet 2020 the same setting gives 0.715 for CMSC against 0.615 for SimCLR 2005.13249 tab. 1. Across label fractions from 0.25 to 1, CMSC stays above SimCLR on both datasets in the table of linear results 2005.13249 app. D.1. The card for this variant is CLOCS: Contrastive Learning of Cardiac Signals Across Space, Time, and Patients.

The method needs patient identity for every training instance, because that is what defines a positive pair 2005.13249 §4.1. It pre-trains on the training splits of PhysioNet 2020 and Chapman, which contain multi-lead data 2005.13249 §5.2. Its baselines are SimCLR, BYOL, MT-SSL, supervised pre-training and random initialisation 2005.13249 §5.4. Results are on ECG only.

Temporal and contextual contrasting, then a linear probe (TS-TCC)

TS-TCC pre-trains a three-block 1-D convolutional encoder on unlabelled series. Each sample gets a weak and a strong augmentation. A small Transformer reads one view up to a cut point and predicts the next latent steps of the other view, and a contrastive term pulls together the two views' summaries of the same sample 2106.14112 §3 2106.14112 §3.2 2106.14112 §3.3. The paper uses 4 Transformer layers, 4 heads, a loss weighting of 1 and 0.7, batch size 128 and 40 epochs, on one RTX 2080 Ti 2106.14112 §4.2. The card for this variant is Time-Series Representation Learning via Temporal and Contextual Contrasting.

With a frozen encoder and one linear layer, the paper reports 90.37 ± 0.34% accuracy on HAR, 83.00 ± 0.71% on Sleep-EDF and 97.23 ± 0.10% on Epilepsy, over five seeds 2106.14112 tab. 2. The supervised row of the same table is 90.14% on HAR, 83.41% on Sleep-EDF and 96.66% on Epilepsy, so the method is ahead on two datasets and behind on Sleep-EDF 2106.14112 tab. 2. Removing the cross-view term costs about five points on HAR, from 87.86% to 82.76% without the contextual term 2106.14112 tab. 4. On cross-domain fault diagnosis the average over twelve transfer scenarios is 67.82%, against 63.83% for supervised training 2106.14112 tab. 3 2106.14112 §5.3.

The paper reports semi-supervised results in text and a figure only: with 1% of labels, about 70% macro F1 on HAR and about 90% on Epilepsy, and 10% of labels comparable to supervised training on all labels 2106.14112 §5.2. The released code applies the augmentations once per dataset and keeps batch normalisation and dropout active during the linear probe, although the paper describes a frozen encoder. Its weak augmentation is scaling only, not the jitter and scaling the paper names. Both are recorded in the card's Paper vs code section.

The HAR data has nine channels. The card's Pitfalls section reports that the strong view as coded can make the channels identical for multichannel input, so check that step before reusing it on multivariate data.

Symbolic word counts over receptive fields (BORF)

BORF slides windows of several sizes, dilations and word lengths over each signal 2311.18029 §4.2 2311.18029 alg. 2. Each window is split into segments, the segment means are z-normalised with the window's own mean and standard deviation, and each mean becomes one of three symbols by Gaussian breakpoints. The symbols form a short word, and the words are counted per channel 2311.18029 §4.1 2311.18029 §3. Computing all segment means once with moving averages means the cost per window no longer grows with window size 2311.18029 §4.1 2311.18029 §4.3.

The count matrix goes through an inverse hyperbolic sine and then a ridge classifier 2311.18029 §5. The method has no random step, so repeated runs give the same features 2311.18029 §5. Each word has one coefficient, and its SHAP value spreads back over the points it came from as a saliency map 2311.18029 §4.5. The paper shows this on one 28-channel EEG case, where choosing which channels to add together moved accuracy from 0.56 to 0.61 2311.18029 §5. That is one case, not a general result.

On 158 UCR and UEA datasets (128 univariate, 30 multivariate) with default splits, the authors report 9th place on average accuracy. BORF is statistically tied with ROCKET and MiniRocket, and only RDST and MR-HYDRA are statistically better 2311.18029 §5 2311.18029 fig. 6. It ranks second on average runtime, tied with HYDRA 2311.18029 §5. The card for this variant is BORF.

The paper sets a flat-window threshold of 0.15, which the released library leaves off by default. The card's Paper vs code section records this and other differences, so set the threshold explicitly when reproducing.

Transformer over observation triplets for sparse clinical series (STraTS)

Each ICU stay is a set of (time, variable, value) triplets. The series is not aggregated onto a regular grid and not imputed 2107.14293 §3.1. Time and value each pass through a learned continuous-value embedding: a small network with one hidden layer of floor(√d) tanh units. The variable name is embedded with a lookup table, and the three parts are summed 2107.14293 §3.2.1. Transformer blocks let every triplet attend to every other, and attention pooling turns the triplets into one vector 2107.14293 §3.2.2 2107.14293 §3.2.3. The vector is joined to a demographics embedding and mapped to a probability by a sigmoid output 2107.14293 §3.2.4 2107.14293 §3.2.5.

Self-supervised pretraining forecasts the variables that are observed in a 2-hour window after an observation window. The whole model is then fine-tuned on the mortality label 2107.14293 §3.2.6. The paper evaluates only the binary task of in-hospital mortality 2107.14293 §3.1. The card for this variant is Self-Supervised Transformer for Sparse and Irregularly Sampled Multivariate Clinical Time-Series.

Transformer over raw per-date reflectance for crop types

Each parcel is one series of 13 Sentinel-2 spectral bands per date, averaged over the parcel's pixels, with no cloud masking or gap filling 1910.10536 §7.2.1 1910.10536 §1. The encoder follows the original Transformer, but it drops the word embedding, because the inputs are already continuous. It adds a positional encoding, applies multi-head self-attention and per-date dense layers with skip connections, and then max-pools over time before a softmax over crop classes 1910.10536 §4.2 1910.10536 §3. The paper compares it with an LSTM, two 1-D CNNs, a random forest and DuPLO on the same parcels 1910.10536 §4.1 1910.10536 §4.3 1910.10536 §4.4 1910.10536 §4.5 1910.10536 §4.6. The card for this variant is Self-attention for raw optical Satellite Time Series Classification.

Each architecture is tuned over 300 hyperparameter sets, and the three best are retrained with their own seeds and reported as mean ± standard deviation 1910.10536 §6 1910.10536 §8.1. On raw data for 23 land-use classes, the Transformer reaches kappa 0.71 and accuracy 0.80, against 0.67 and 0.79 for TempCNN and 0.71 and 0.81 for the LSTM 1910.10536 tab. 1. On the 12 land-cover classes the Transformer reaches 0.89 accuracy on raw data, against 0.83 for TempCNN and 0.90 for the LSTM 1910.10536 tab. 1. The authors report that preprocessing raised every model's scores, and that the choice of model mattered less on preprocessed data 1910.10536 §8.1. The random forest, with spectral indices, reaches accuracy 0.71 on raw 23-class data 1910.10536 tab. 1.

The paper reads a gradient-based saliency check as the Transformer and the LSTM suppressing cloudy dates, while the convolutional models react to some of them. Attention maps from the first of three layers point to a few observations 1910.10536 §8.3 1910.10536 §8.4. That is one model run and one example per class, and the authors do not test it on other data. Two details of the training setup affect reuse. The paper stops training on a rule over the last ten epochs without saying which partition the rule reads 1910.10536 §5. The card reports that the released code reads the test partition, and that it does not use the tuned base learning rate for the Transformer (see the card's Paper vs code section).

The results are author-reported. On MIMIC-III (24-hour mortality) and PhysioNet-2012 (48-hour mortality), the paper's model has the best ROC-AUC, PR-AUC and min(recall, precision) against six baselines 2107.14293 §4.5 2107.14293 tab. 4. On PR-AUC it scores 0.577 against 0.559 for the best baseline on MIMIC-III, and 0.498 against 0.481 on PhysioNet-2012. The authors report these as gains of 3.2% and 3.5% 2107.14293 §4.5. The ROC-AUC gaps are smaller: 0.891 against 0.886 on MIMIC-III 2107.14293 tab. 4. Without self-supervision the PR-AUC falls to 0.546 on MIMIC-III and 0.467 on PhysioNet-2012 2107.14293 tab. 5. The interpretable variant, I-STraTS, gives a contribution score per observation and per demographic, but the paper presents one case study rather than a validated attribution method 2107.14293 §3.3 2107.14293 §4.7.

Image encodings of the series with a Tiled CNN (GASF, GADF, MTF)

The series is turned into a square image, so that an image model can read it. A Gramian angular field maps each rescaled value to an angle and stores the cosines of angle sums (GASF) or the sines of angle differences (GADF) 1506.00327 §2 1506.00327 §2.1. A Markov transition field counts transitions between quantile bins and places each probability at a pair of time steps 1506.00327 §2.2. The three images are stacked as channels at one size S 1506.00327 §3.3. Two Tiled CNN layers, pretrained with topographic ICA, feed a linear SVM, and S, the quantile count Q and the SVM penalty are chosen per dataset by lowest training-set error 1506.00327 §3.2 1506.00327 fig. 3. The unlabelled pretraining uses all series, test inputs included 1506.00327 §3.2.

On 20 UCR datasets, each with its own train and test split, the paper's table has this variant at the lowest error on 9 of them 1506.00327 tab. 1. The competitor errors are copied from earlier publications, not rerun, and each entry is one run with no variance 1506.00327 §3. The card for this variant is Imaging Time-Series to Improve Classification and Imputation.

The same encoding is also used for imputation. A denoising autoencoder with one 500-unit sigmoid layer, trained on GASF images of five UCR series with 20% of points zeroed, lowers the error over missing points by 12.18% to 48.02% against the same autoencoder on raw series 1506.00327 §4.1 1506.00327 §4.2 1506.00327 tab. 2. Over the whole recovered series, the GASF route has a higher error in every row of that table 1506.00327 tab. 2. The inverse from the diagonal to values holds only for series rescaled to [0, 1] 1506.00327 §4. See time series imputation for the task itself.

Classifiers on single-grasp tactile series (SVM, LSTM, k-NN)

Each squeeze of a soft object gives a short multi-channel series. The channels depend on the device: motor position and current, gap position and current, or pressure cells on the Barrett Hand 2204.06343 §3.2. The same series are classified three ways. An SVM reads nine statistics per channel, an LSTM reads the raw series padded with zeros, and k-NN reads either form 2204.06343 §4.1 2204.06343 §4.2. The SVM grid covers linear, sigmoid and RBF kernels 2204.06343 §4.2.2, and the LSTM grid has 336 trained models 2204.06343 §4.2.3. The card for this variant is Single-grasp Deformable Object Discrimination.

On the 20 polyurethane foams, the Barrett Hand reaches 94–99% with SVM or LSTM across actions 2204.06343 §5.2 2204.06343 fig. 5. The Robotiq 2F-85 reaches 68–77% with SVM, and the RG6 reaches over 75% with LSTM 2204.06343 §5.2. The text gives no number for k-NN on foams and says only that it lags 2204.06343 §5.2. On the ordinary object set, the Barrett and Robotiq are close to perfect 2204.06343 §5.1.

The paper also asks what a model learns when the setting changes. Transfer across finger configurations fails for the SoftHand and the Barrett Hand 2204.06343 §5.4 2204.06343 fig. 6. On the Barrett foams, tactile cells alone come close to the full sensor set, and on the parallel grippers the gap position alone is close to gap position plus current 2204.06343 §5.5 2204.06343 fig. 7.

The paper does not say explicitly that the reported accuracy is on the held-out test split, and it selects settings on the validation split 2204.06343 §4.2.2 2204.06343 §4.2.3. Split proportions are not stated 2204.06343 §4.2.4, and the side on which the LSTM padding goes is not given 2204.06343 §4.2.3.

CNN on robot joint and force series for task labels

A sample is a 1000 × 24 matrix: 10 s of joint positions, velocities, efforts and wrist force/torque at 100 Hz. Shorter recordings are zero-padded to the same length 2402.10220 §IV-A 2402.10220 §IV-E 2402.10220 §III-D. The classifier is a stack of convolution and pooling blocks ending in a softmax over six tasks 2402.10220 fig. 7a. Two parallel-stream variants, called dual and quadruplet, concatenate their branches 2402.10220 fig. 7b 2402.10220 fig. 7c. Training uses categorical cross-entropy 2402.10220 eq. 4, with random splits of 80/10/10, 70/15/15 and 60/20/20 2402.10220 §IV-B. The card for this variant is Towards Semi-Autonomous Robotic Arm Manipulation Operator Intention Detection from Force Data.

The author-reported accuracy on the glovebox set is 86% for the single stream, 84% for dual and 89% for quadruplet 2402.10220 §IV-C 2402.10220 fig. 12. On the open Kinova benchmark the CNN reaches 100% ± 2% F1 across splits 2402.10220 §IV-B 2402.10220 fig. 9. The only baseline is an LSTM on the spatial mean of the channels 2402.10220 §IV-E.1 2402.10220 tab. III. The same paper also forecasts the next 5 s of readings, which is a forecasting task, covered on the forecasting page.

The data are small: about 20 to 40 samples per class on the open benchmark and 20 per task on the glovebox set 2402.10220 tab. I 2402.10220 §III-D. The abstract calls the 89% an F1 score, but §IV-C calls it accuracy 2402.10220 p. 1 2402.10220 §IV-C. Compare it only with a number that uses the same metric.

Random Forest on daily multispectral series for burned areas

Each 500 m grid cell gets a daily series for each of four bands (blue, green, red and near-infrared) from the Brazilian WFI sensors. A day without an image takes the last earlier value 2609.23687 §2.2.11. For one year at a time, these daily values form the classification input [arxiv:2609.23687 §2.2.15]. A scikit-learn Random Forest is fitted, with the number of trees and the maximum depth chosen by grid search 2609.23687 §2.2.16. Each cell is labelled by hand as totally burned (TB), partially burned (PB) or non-burned (NB) 2609.23687 §2.2.13. Three relabelling schemes are tried: TB as burned, TB and PB as burned, and three separate classes 2609.23687 tab. 5. The card for this variant is A multi-temporal dataset for mapping burned areas in the Brazilian Cerrado using time series of remote sensing imagery.

On held-out cells, the TB-as-burned scheme has an F1 of 0.94, 0.95 and 0.92 for 2020 to 2022, and the three-class scheme has 0.79, 0.79 and 0.74 2609.23687 tab. 7. These are means over 200 random runs 2609.23687 §2.2.16. The authors' fifth validation setup, a three-class model read as TB-as-burned, has a burned-area IoU of 0.66, 0.71 and 0.49 2609.23687 tab. 8. Overall IoU there is 0.98, 0.98 and 0.95, but non-burned cells make up 86% to 88% of the grid, so the overall figure says little about the burned class 2609.23687 §3.3.

The split is a random 30% of cells for testing, and the paper gives no separate check for spatial correlation between neighbouring cells 2609.23687 §2.2.17. The random-forest settings that matter for reproduction, such as the split criterion and the seed, are not stated 2609.23687 §2.2.16.

Benchmark evidence

Bake Off Redux takes the best algorithm from each of eight families and runs them against 1-NN DTW on 112 equal-length UCR problems, with 30 resamples per problem 2304.13029 §3 2304.13029 §5. It adds 30 new datasets to check that the winners were not tuned only to the archive 2304.13029 §3.1 2304.13029 §5. Accuracy and training time from its results table:

2304.13029 tab. 14 2304.13029 tab. 19

• On the 112 problems, HC2 and MR-Hydra are first and second on accuracy. On the 30 new datasets, the top accuracy clique contains MR-Hydra, HC2, RDST, QUANT and FreshPRINCE 2304.13029 §5. On all 142 datasets, HC2 and MR-Hydra share the top clique 2304.13029 §5.
• The authors suggest HC2 and MR-Hydra as a default starting point when there is no prior knowledge about the data 2304.13029 §7. HC2 is the method introduced in its own card, which reports it as significantly more accurate than four earlier methods on UCR problems 2104.07551 §5.1.
• QUANT is the fastest by a wide margin, and its accuracy is lower than the top two's 2304.13029 tab. 14 2304.13029 tab. 19.
• The paper compares univariate series only 2304.13029 §3. The 112 problems were used to support many proposed methods, so the authors worry that design choices were tuned to them 2304.13029 §5 2304.13029 §7.
• It uses 30 resamples where the 2017 bake off used 100, so small accuracy differences may reorder close methods 2304.13029 §3.

The card for this benchmark is Bake off redux: a review and experimental evaluation of recent time series classification algorithms.

When to use which

The CNN card calls labelled data scarce and names it as the main obstacle to better results 1804.02386 §6. This is the authors' view of their own problem. It does not show how the methods compare at other sizes.

The clinical Transformer card notes that the released code departs from the paper in several places, for example attention scaling, residual connections, the activation function and the pretraining population. Expect the released scripts to give different numbers from the paper's table, and read the card's Paper vs code section before implementing 2107.14293 §3.2.2 (see the card).

Implementation sketch

The sketch below follows the random-kernel recipe. It is our own simplification. The paper's kernel count, dilations, padding and biases are set in its code, and the card says which choices differ from the text 2102.00457 §3.2.

# Random-kernel classifier for univariate series  [arxiv:2102.00457 §3]
# Our sketch; dilations, padding and bias details are left out
def fit(train_series, labels):
    # Two views of each series: the raw values and their first difference
    # [arxiv:2102.00457 §3.1]
    views = [(s, diff(s)) for s in train_series]
    # Fixed kernels; dilations and biases are fitted from the training data
    # [arxiv:2102.00457 §3.2]
    params = fit_dilations_and_biases(train_series)
    # Four pooled values per kernel, dilation and bias  [arxiv:2102.00457 §3.4]
    features = [pool_all(view, params) for view in views]
    # Linear classifier on the pooled features  [arxiv:2102.00457 §3.5]
    return ridge_fit(features, labels)

For the CNN, the same outer shape holds, with the features computed by convolution layers and a softmax over the classes instead of a linear classifier 1804.02386 §3.2 1804.02386 §3.2.5.

Contradictions

Within one source:

• The CNN paper states that epochs are chosen on a validation score in one place and on the test curve in another 1804.02386 §3.2.8 1804.02386 §5.1. The card treats the test-based choice as the one used, so its 82.3% and 84.8% figures may be optimistic.
• Table 5 of the CNN paper sums to 84.3% on the diagonal, while Table 4 reports 84.8% for the same ensemble 1804.02386 tab. 5 1804.02386 tab. 4.
• The MultiRocket paper gives three different speed ratios against MiniRocket: about ten times 2102.00457 §1, about four times 2102.00457 §4.2, and about three times from its own Table 2 totals 2102.00457 tab. 2. The timings measure different things, so the ratios should not be mixed.
• The tutorial recommends training large models rather than growing small ones, and also recommends pre-trained models for small datasets 2205.01138 §5.4.1 2205.01138 §5.4.2. It does not say which applies to a given series.
• The imaging paper's text gives a search grid of S in {16, 24, 32, 40, 48} and Q in {8, 16, 32, 64} 1506.00327 §3.2, but Table 1 reports values outside it: S = 8 for FaceAll and S = 64 for Beef, and Q = 40 for Beef and Q = 48 for Coffee 1506.00327 tab. 1. The grid or the table is wrong. This page does not decide which, and the card's Pitfalls section suggests the table's values for reproduction.
• The CLOCS paper says all three of its methods beat SimCLR in every linear experiment 2005.13249 §6.2. Its own appendix table has a case where CMLC is slightly below SimCLR: 0.580 against 0.585 on PhysioNet 2020 at E=128, F=0.25 2005.13249 app. D.1. The authors' claim is stronger than their table.
• The TS-TCC text says it improves by at least 7 points in seven of its eight winning transfer scenarios, except D→B. Table 3 shows only five of the seven non-D→B wins above 7 points, since A→C (+6.56) and B→C (+6.71) fall short 2106.14112 §5.3 2106.14112 tab. 3. The average gain of about 4 points does match the table.
• The benchmark's text says HC2 is about 0.5% more accurate than MR-Hydra and more than 6% more accurate than Proximity Forest 2304.13029 §5. Its results table gives 0.891 against 0.884 and 0.891 against 0.837, which are 0.7 and 5.4 points 2304.13029 tab. 14. This page uses the table.
• The glovebox paper calls its 89% result an F1 score in the abstract and §V-A, and accuracy in §IV-C 2402.10220 p. 1 2402.10220 §V-A 2402.10220 §IV-C. Its experiment E4 is 97% in the text and 100% in its figure 2402.10220 §V-C 2402.10220 fig. 10. This page reports the numbers as the paper gives them and does not pick one.
• The HC2 paper's text and its tables disagree on which large-dataset runs finished on FruitFlies and InsectSound, and whether one of the methods ran out of memory 2104.07551 §7 2104.07551 tab. 10 2104.07551 tab. 11. This page does not use those large-dataset numbers.
• The burned-area paper's abstract gives burned-area omission as 22%, 28% and 59%, and commission as 46%, 43% and 46% 2609.23687 p. 1. Its Table 11 gives commission of 0.227, 0.282 and 0.593 and omission of 0.461, 0.432 and 0.460 for the same approach, so the two labels appear to be swapped 2609.23687 tab. 11. The paper's §3.3 IoU values also differ from Table 8 in several rows 2609.23687 §3.3 2609.23687 tab. 8. This page uses the tables.
• The benchmark's critical difference diagrams can rank methods differently from their pairwise win counts. The authors give a case where a method wins more pairwise comparisons yet ranks lower in the diagram 2304.13029 §5. The authors' own comparison matrices should be checked before ranking.

Across sources:

• The thirteen cards with experiments use different tasks and protocols: travel-mode labels from GPS with a random segment split, UCR and UEA archive datasets with resampling 2104.07551 §4, crop types from satellite series over three Bavarian regions with held-out blocks 1910.10536 §7, ECG arrhythmia labels with patient-level splits, HAR, Sleep-EDF and Epilepsy with the TS-TCC splits 2106.14112 §4.1, ICU mortality with a patient-level split on MIMIC-III 2107.14293 §4.1, the benchmark's resampled UCR and new datasets 2304.13029 §3, BORF's 158 archive datasets with default splits 2311.18029 §5, the imaging paper's 20 UCR datasets with copied baselines 1506.00327 §3, and the two robot studies, with one data set per device 2204.06343 §5 and random splits over small glovebox sets 2402.10220 §IV-B, and the burned-area cells, split at random 70/30 and repeated 200 times 2609.23687 §2.2.16. Their accuracies and AUCs cannot be ranked against each other. The tutorial runs no experiments.
• The benchmark and the MultiRocket paper both put HC2 at or near the top on UCR-style data. They use different dataset counts (112 and 109) and different protocols, so their numbers should not be combined 2304.13029 §3 2102.00457 §4.1.
• HC2 timings disagree across sources. The HC2 paper gives about 340 hours for one sequential resample of 112 UCR problems 2104.07551 tab. 4. The benchmark gives 263.89 hours for HC2 over 142 problems, with its own protocol 2304.13029 tab. 19. The two totals measure different runs, and the timing methods differ, so do not compare them directly.
• The HC2 paper and the benchmark both report 112 UCR problems for HC2's accuracy. The cards do not show that the problem sets are identical, so the two accuracy results should not be merged 2104.07551 §4 2304.13029 §3.
• The benchmark and BORF both rank MR-HYDRA among the strongest methods, but BORF is not in the benchmark, and BORF's comparisons use 158 datasets with default splits. Their rankings should not be combined 2311.18029 §5 2304.13029 §3.

Papers

• Imaging Time-Series to Improve Classification and Imputation (Wang and Oates, arXiv:1506.00327): GASF, GADF and MTF image encodings with a Tiled CNN, with author-reported lowest error on 9 of 20 UCR datasets, and a denoising autoencoder that lowers error on missing points.
• Self-attention for raw optical Satellite Time Series Classification (Rußwurm and Körner, arXiv:1910.10536): a Transformer encoder on raw Sentinel-2 per-date reflectance for crop types, with author-reported kappa and accuracy against an LSTM, two CNNs and a random forest.
• Inferring transportation modes from GPS trajectories using a convolutional neural network: a 1-D CNN over four per-point motion channels for travel-mode classification, with an ensemble reported at 84.8% on GeoLife.
• CLOCS: Contrastive Learning of Cardiac Signals Across Space, Time, and Patients (Kiyasseh et al., arXiv:2005.13249): patient-level contrastive pre-training of a 1-D CNN on ECG, with author-reported linear and fine-tuning AUCs on four ECG datasets.
• HIVE-COTE 2.0: a new meta ensemble for time series classification (Middlehurst et al., arXiv:2104.07551): a meta ensemble of four modules, with author-reported accuracy on 112 UCR and 26 UEA problems and a 340-hour sequential run time for one UCR resample.
• MultiRocket (Tan et al., arXiv:2102.00457): fixed random kernels on the series and its difference, four pooling operators and a ridge classifier, with author-reported UCR results.
• Time-Series Representation Learning via Temporal and Contextual Contrasting (Eldele et al., arXiv:2106.14112): unsupervised temporal and contextual contrasting of augmented views, with author-reported linear-probe accuracy on HAR, Sleep-EDF and Epilepsy.
• Self-Supervised Transformer for Sparse and Irregularly Sampled Multivariate Clinical Time-Series (Tipirneni and Reddy, arXiv:2107.14293): a Transformer over observation triplets with forecasting pretraining, with author-reported gains in PR-AUC on MIMIC-III and PhysioNet-2012 mortality.
• Single-grasp Deformable Object Discrimination: the Effect of Gripper Morphology, Sensing Modalities, and Action Parameters (Pliska et al., arXiv:2204.06343): SVM, LSTM and k-NN classifiers on squeeze series from four robot hands and grippers, with author-reported accuracy on 20 foams.
• Transformers in Time-series Analysis: A Tutorial: a tutorial on Transformer blocks and their time-series adaptations, including a classification head, with no experiments of its own.
• Bake off redux: a review and experimental evaluation of recent time series classification algorithms (Middlehurst et al., arXiv:2304.13029): a benchmark of eight algorithm families on 112 UCR problems and 30 new datasets, with HC2 and MR-Hydra the most accurate.
• BORF (Spinnato et al., arXiv:2311.18029): deterministic symbolic word counts over receptive fields with a ridge classifier, with author-reported accuracy statistically tied with ROCKET and MiniRocket on 158 datasets.
• Towards Semi-Autonomous Robotic Arm Manipulation Operator Intention Detection from Force Data (Alharthi et al., arXiv:2402.10220): a CNN over 24 channels of robot joint and force readings, classifying six glovebox tasks, with author-reported accuracy and F1.
• A multi-temporal dataset for mapping burned areas in the Brazilian Cerrado using time series of remote sensing imagery (de Oliveira and Körting, arXiv:2609.23687): a labelled daily satellite dataset for one park, with a Random Forest baseline that reaches F1 0.92 to 0.95 on the binary burned label, author-reported.