Do signature-based features improve human activity recognition over raw-signal baselines?

Three representations, one leakage-aware protocol

The study asks whether signature-based feature learning improves human activity recognition (HAR) relative to raw-signal baselines, and how that answer changes with embedding strategy, truncation depth, model choice, and sensor configuration. Experiments used the UCI HAR dataset with the original train–test split preserved and subject-disjoint validation to prevent leakage, so reported gains reflect generalisation rather than optimistic bias from overlapping subjects.

Three representations were compared under identical preprocessing and validation: raw flattened signals, time-augmented paths, and lead–lag transformed paths. Signature features were computed at multiple truncation depths and evaluated with a multilayer perceptron (MLP) and a Random Forest (RF). A prior K-means-based feature reduction study was also reproduced for comparison. Because all configurations shared the same partitions, preprocessing, and metrics, differences between them can be attributed to representation and classifier rather than to pipeline variation.

Signatures make temporal order explicit before classification

Raw windows preserve original sensor measurements but do not explicitly encode temporal ordering or cross-channel interactions, leaving the classifier to infer those patterns implicitly. Signature transforms instead shift part of that burden to the feature construction stage, providing a structured representation of temporal interactions in which higher-order terms capture increasingly complex dependencies across time. For multivariate signals, the complexity of this representation is controlled explicitly through the truncation depth.

The two embeddings expose different structure. Time augmentation introduces a strictly increasing temporal dimension, preserving ordering. Lead–lag embedding aligns delayed and current values, exposing inter-channel interactions at lower orders. The paper's interpretation is that this is why lead–lag can be competitive at moderate depths, while time augmentation gives the highest overall performance: for fixed-window HAR data, preserving temporal order with limited dimensional expansion balances expressiveness and tractability better than the more complex lead–lag embedding.

Random Forest gains, and depth keeps helping

With Random Forest, signature features produced clear improvements. The best configuration was time-augmented six-channel signatures at depth 6 using an entropy-based RF, achieving 0.858 accuracy and 0.859 macro F1, which outperformed the strongest raw baseline at 0.816 accuracy. For RF models, increasing truncation depth consistently improved results, indicating that higher-order interactions remained informative for this classifier.

The paper attributes this to the RF's ability to exploit structured, high-order summaries of temporal interactions without requiring complex optimisation. The broader claim the authors draw is that signature-based representations are particularly effective when representation learning is decoupled from model training rather than embedded within end-to-end optimisation. That is an interpretation of the pattern, not a separately measured comparison.

MLP models never beat the raw baseline

The classifier dependence runs in both directions. For MLP models, raw flattened inputs outperformed all signature-based configurations, and MLP models did not exceed raw baselines. The authors suggest the network effectively exploits the structured fixed-length input on its own, so pre-computing signature features adds dimensionality without adding usable information.

MLP performance was also less stable across truncation depths, with improvements plateauing or declining at higher depths. The paper frames this as a trade-off between representational richness and learnability: deeper signatures increase expressiveness but also introduce high dimensionality that can hinder optimisation. This is why truncation depth should be treated as a key modelling hyperparameter rather than a technical detail, and why the same depth that helps RF can hurt MLP.

Six channels beat accelerometer-only inputs

Sensor configuration mattered as well. Accelerometer-only data can approximate full inertial performance to a considerable extent, particularly in MLP settings where lower dimensionality aids optimisation. However, the best overall results were obtained using six-channel inputs, indicating that gyroscope signals provide complementary information that improves classification.

This result sits alongside the representation and classifier findings rather than replacing them: the best configuration combined time-augmented six-channel signatures with an RF classifier. Readers comparing sensor setups should note that the accelerometer-only approximation is described as considerable but not equal to the six-channel result, and that the advantage appears in the paper's own fixed-window UCI HAR setting.

Temporal structure beats feature reduction alone

The closest prior study on UCI HAR used K-means-based feature reduction over the 561 handcrafted features followed by Naive Bayes classification. It reported 76.77% accuracy using all features and a maximum of 83.35% at k = 70. The present study's reproduction confirmed the general trend of improvement with reduction but reached lower performance, 80.10% at k = 70, a discrepancy the authors attribute to incomplete reporting of preprocessing, parameter settings, and validation procedures in the original work.

The comparison clarifies what the signature framework contributes. The K-means approach operates on fixed handcrafted features and focuses on redundancy reduction within a static representation. Signature-based representations instead address how temporal structure should be represented before classification, explicitly encoding ordering and inter-channel interactions. The best model in this study reached 85.816% on a held-out test set, above both the reproduced 80.10% and the originally reported 83.35%. The authors read this as evidence that gains come from encoding temporal dynamics rather than from dimensionality optimisation.

Single dataset, two model families, high-dimensional features

Several limits qualify how far these results travel. Experiments were restricted to a single dataset, which may limit generalisability to other domains or sensor configurations. Signature features can lead to high-dimensional representations, particularly at larger truncation depths, which may introduce computational challenges and affect model stability. Only two model families were considered, so additional architectures such as convolutional or recurrent models were not explored and could change the relative performance of representations.

The reproducibility framework is a strength rather than a caveat: a pre-specified protocol, public code, a model card for the best-performing model, and a Docker environment support independent verification, and the leakage-aware subject-disjoint validation strengthens confidence that the reported gains are genuine. The practical reading is that the paper's central variable is the representation–classifier pairing, and its evidence is bounded to the UCI HAR dataset and the MLP and RF families tested there.

Choose the classifier before trusting the signature

The paper's most decision-relevant result is not that signatures help HAR, but that they help only in specific pairings. If the classifier is a Random Forest, time-augmented six-channel signatures at depth 6 with entropy-based RF is the configuration that produced the best measured outcome, 0.858 accuracy and 0.859 macro F1 against a 0.816 raw baseline. If the classifier is an MLP, the raw flattened representation remained better in every signature configuration tested.

Truncation depth should be tuned per classifier rather than fixed in advance, since RF improved monotonically with depth while MLP plateaued or declined. And when comparing against feature-reduction approaches, the relevant contrast is representation design versus redundancy reduction: the reproduced K-means pipeline peaked at 80.10% at k = 70, while the best signature model reached 85.816% on the held-out test set. The priority this paper establishes is to treat representation and classifier as a joint decision, and to verify that pairing under leakage-aware validation before assuming signatures will transfer.

Sources used in this answer