CoChat

Ensembles of Deep LSTM Learners for Activity Recognition using Wearables

Open access
Guan, Yu, Plötz, Thomas · 2017 · Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies

The DOI matches its Crossref record; 53 of the 65 references it cites were found in a registry or an index, and 10 are not checked yet; none of the 62 sources its claims cite is checked for support yet.

Abstract

Recently, deep learning (DL) methods have been introduced very successfully into human activity recognition (HAR) scenarios in ubiquitous and wearable computing. Especially the prospect of overcoming the need for manual feature design combined with superior classification capabilities render deep neural networks very attractive for real-life HAR applications. Even though DL-based approaches now outperform the state-of-the-art in a number of recognition tasks, still substantial challenges remain. Most prominently, issues with real-life datasets, typically including imbalanced datasets and problematic data quality, still limit the effectiveness of activity recognition using wearables. In this paper we tackle such challenges through Ensembles of deep Long Short Term Memory (LSTM) networks. LSTM networks currently represent the state-of-the-art with superior classification performance on relevant HAR benchmark datasets. We have developed modified training procedures for LSTM networks and combine sets of diverse LSTM learners into classifier collectives. We demonstrate that Ensembles of deep LSTM learners outperform individual LSTM networks and thus push the state-of-the-art in human activity recognition using wearables. Through an extensive experimental evaluation on three standard benchmarks (Opportunity, PAMAP2, Skoda) we demonstrate the excellent recognition capabilities of our approach and its potential for real-life applications of human activity recognition.

Checked 2026-10-05 · 10 references not checked yet

Overview

CoChat Summary

TL;DR

The paper introduces ensembles of deep LSTM learners for wearable activity recognition, using an epoch-wise bagging-style training procedure and score-level fusion of multiple LSTM models. Across Opportunity, PAMAP2, and Skoda, the ensembles outperform individual LSTMs and state-of-the-art baselines, with the largest gains on the challenging, imbalanced Opportunity dataset.

Key findings

  • Ensembles of LSTM learners improve over single LSTMs on all three benchmarks. For Opportunity, single CE LSTM F1 = 0.674 ± 0.021 vs. Ensemble M=10 CE = 0.716 ± 0.011 and M=20 CE = 0.720 ± 0.011; the best combined CE+F1 ensemble reaches 0.726 ± 0.008. For PAMAP2, single CE = 0.802 ± 0.045 vs. Ensemble M=10 CE = 0.848 ± 0.021 and M=20 CE = 0.852 ± 0.026; combined CE+F1 = 0.854 ± 0.026. For Skoda, single CE = 0.918 ± 0.007 vs. Ensemble M=10 CE = 0.925 ± 0.002 and M=20 CE = 0.926 ± 0.001; combined CE+F1 = 0.924 ± 0.002.
  • The LSTM ensemble outperforms reported state-of-the-art baselines: LSTM baseline 0.659/0.756/0.904, DeepConvLSTM 0.672/0.748/0.912, and LSTM-S 0.684/0.838/0.921 for Opportunity/PAMAP2/Skoda, respectively, versus the ensemble’s 0.726/0.854/0.924.
  • Performance generally plateaus at M=10 base learners; increasing to M=20 gives no statistically significant improvement except for Opportunity with F1 loss, where the improvement is “just significant.” Combining CE and F1 loss functions further improves Opportunity but not PAMAP2, and Skoda shows no substantial boost.
  • The training procedure intentionally omits data: on average about 35%–38% of training data is not used per epoch when using a uniformly distributed random number generator for training frames.

Methods

  • The authors use two-layer LSTM networks with 256 LSTM units per layer, dropout 0.5 at both hidden layers, ADAM with learning rate 0.001, implemented in Lasagne/Theano and trained on an NVIDIA Tesla K40. Training uses randomly drawn mini-batch sizes from U(128, 256) and frame lengths from U(16, 32); test-time inference is sample-wise with mini-batch size and frame length set to 1.
  • The core method is epoch-wise bagging: for each epoch, random mini-batch size, random starting positions, and random frame lengths are drawn; one LSTM learner is produced per epoch (MaxEpoch = 100). The M best models on a validation set are retained and fused by arithmetic mean of their softmax class-probability vectors. Both cross-entropy and F1 loss are explored.
  • Evaluation uses three standard benchmarks: Opportunity, PAMAP2, and Skoda, with hold-out validation/test protocols replicating prior work. Results are mean F1 scores over 30 repetitions, with two-tailed independent t-tests for significance.

Limitations

  • The paper states that implementation issues for field deployments or products are not yet addressed; ensembles inflate the parameter space, though the authors argue pruning remains accessible because the base LSTM models are unchanged.
  • The paper notes that F1 loss is a global measure and cannot directly measure individual sample losses; because LSTMs are trained on mini-batches, it may not precisely approximate the true gradient toward optimal F1 over the whole training set.
  • The paper states that Skoda shows no substantial performance boost, attributed to only one worker being considered and the task lacking challenging test data from different environments; the baseline models already achieve high specialisation (F1 > 0.9). Evident from the text: evaluation is limited to three benchmark datasets and does not include an in-the-wild deployment, and frame-based results from prior work were converted to sample-wise predictions for comparability.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3090076

Cites

53 of 63 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

0 Valid · 11 Not valid · 48 Not checked · 80 Own results and interpretations

55 more claims

Create a free account to see all 139 claims and to verify the 62 references still open.

See all claims — free account
  • It may state two separate claims as one. 32
  • The quote only partly states this claim. 2
  • Probably not a substantial claim. 59

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Research works printed but not linked: 2

  1. James L McClelland and David E Rumelhart . 1986. Parallel distributed processing . MIT Press . James L McClelland and David E Rumelhart. 1986. Parallel distributed processing. MIT Press.
  2. Theano Development Team . 2016 . Theano: A Python framework for fast computation of mathematical expressions. arXiv e-prints abs/1605.02688 (May 2016). http://arxiv.org/abs/1605.02688 Theano Development Team. 2016. Theano: A Python framework for fast computation of mathematical expressions. arXiv e-prints abs/1605.02688 (May 2016). http://arxiv.org/abs/1605.02688 · arXiv

Cited by

See all 539

Metadata

Authors
Guan, Y., & Plötz, T.
Affiliations
  • Newcastle University (GB)
  • Georgia Institute of Technology (US)
Venue
Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies
Published
2017-06-30
Volume
1
Issue
2
Pages
1-28
ISSN
2474-9567 (ISSN-L)
OpenAlex
W2604630936
Cited by
511
References
65
Citations by year
2026: 33 · 2025: 35 · 2024: 59 · 2023: 60 · 2022: 70 · 2021: 82 · 2020: 87 · 2019: 46 · 2018: 33 · 2017: 6
FWCI
19.53
Type
journal article
Language
en
Open access
Locations
  • Proceedings of the ACM on Interactive Mobile Wearable and Ubiquitous Technologies (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
Engineering and Physical Sciences Research Council (EP/M020576/1)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Context-Aware Activity Recognition Systems
  • Physical Sciences › Computer Science › Human-Computer Interaction › Hand Gesture Recognition Systems
  • Physical Sciences › Engineering › Biomedical Engineering › Gait Recognition and Analysis
Keywords
human activity recognition, wearable sensors, deep learning, ensemble classifiers, class imbalance, data quality, training protocols, PAMAP2, Opportunity dataset

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-10-05, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-05 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.