Pith. sign in

REVIEW 4 major objections 5 minor 66 references

ADLGen: Synthesizing Symbolic, Event-Triggered Sensor Sequences for Human Activity Modeling

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ADLGen claims synthetic smart-home sensor logs can match real data in statistics and semantics, and that augmenting scarce real samples with them lifts rare-activity recognition macro F1 from 0.76 to 0.99.

desk verdict Promising generative framework for ADL sensor streams, but the evaluation is compromised by an internal contradiction and self-referential LLM scoring; needs major revision, not acceptance. read the letter →

arxiv 2505.17987 v1 pith:FNUN6EXB submitted 2025-05-23 cs.LG cs.AI

Weihang You , Hanqi Jiang , Zishuai Liu , Zihang Xie , Tianming Liu , Jin Lu , Fei Dou This is my paper · ORCID
classification cs.LGcs.AI
keywords ADLsynthesissensoreventsequencesactivityrecognitiondataaugmentationtransformerLLMrefinementsmarthomerareclasses
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

ADLGen is a generative framework for synthesizing sensor event sequences of Activities of Daily Living in smart homes. The paper argues that a decoder-only Transformer with sign-based tokenization, symbolic-temporal decoupling, and an LLM-based generate-evaluate-refine loop produces sequences that match the statistics of real data, satisfy physical plausibility constraints, and score higher on LLM-judged semantic quality than real examples. The practical payoff is data augmentation: with only 15 real samples of a rare activity, adding generated sequences raises macro F1 from 0.76 to 0.99, and models trained purely on synthetic data nearly match models trained on real data. If correct, the framework offers a privacy-preserving route to training activity-recognition systems without extensive real-world sensor deployment.

What carries the argument

The load-bearing object is the sign-based, symbolic-temporal decoupled representation and the LLM refinement loop. Each sensor event is written as a token whose sign encodes the binary state, so +M001 means ON/OPEN and -M001 means OFF/CLOSED, halving vocabulary size and sequence length while preserving sensor identity. Sensor tokens and timestamps are fed as two parallel streams and fused through $E = E_{sens} + \omega_p PE + \omega_t E_{temp}$, keeping the logical grammar of sensor firings contiguous. During inference, a context-aware temperature schedule and a floorplan adjacency matrix $A_{map}$ enforce physical plausibility. An LLM then translates sequences to natural-language descriptions, scores them on a three-level hierarchy of fundamental logic, behavioral coherence, and temporal consistency, and emits executable INSERT, DELETE, and REORDER rules that are applied programmatically.

What would settle it

Collect independent human judgments of behavioral plausibility for ADLGen, baseline, and real sequences; if human ratings do not rank them the same way as the LLM metric, the semantic-quality result is an artifact of the metric. A second falsifier is to apply the full pipeline to a different dataset with a different floorplan and check whether the validity rate stays near 100% without retuning; if physically impossible transitions reappear, the layout-aware component is overfit to the original home.

Watch

Extended reading notes

Core claim

On the paper's own terms, ADLGen establishes that event-triggered, symbolic ADL sensor streams can be generated with both statistical fidelity and semantic coherence. The full pipeline achieves an MMD2 of 0.0019 versus 0.0088 for the best baseline, 100% validity under the floorplan adjacency check, and an LLM-assessed semantic score of 4.67 out of 5, higher than the raw data's 4.23. In downstream recognition, training on ADLGen synthetic data and testing on real data gives MacroAcc 79.12 and MacroF1 72.90, close to the real-data benchmark of 79.78 and 79.73. The authors attribute the gains to three design choices: sign-based tokenization that factorizes sensor identity and binary state, symbolic-temporal decoupling that keeps the sensor grammar contiguous, and a two-tier LLM evaluation that generates insertion, deletion, and reordering rules to correct violations.

Load-bearing premise

The LLM's semantic-quality score, produced by the same model that writes the corrections, is a valid measure of real-world behavioral authenticity, and the generated refinement rules do not introduce errors that the metric cannot detect.

Editorial extensions

If this is right

  • Training an activity recognizer on ADLGen synthetic data and testing on real data nearly matches training on real data: MacroAcc 79.12 versus 79.78 and MacroF1 72.90 versus 79.73.
  • Augmenting 15 real samples of rare activities with synthetic sequences raises MacroF1 from 0.76 to 0.99, and similar large gains appear at the 30-sample regime.
  • Generated sequences pass 100% of floorplan adjacency validity checks, so the pipeline avoids physically impossible sensor transitions.
  • LLM-evaluated semantic quality reaches 4.67 out of 5, exceeding the raw real data's 4.23, suggesting refinement produces behaviorally richer sequences than the original captures.
  • Cross-floorplan transfer improves: augmenting the source home's data with ADLGen raises accuracy and F1 on a different home's data by roughly 6 to 7 percentage points under the TDOST protocol.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: because the semantic-quality metric and the refinement rules come from the same LLM, the 4.67 score exceeding real data's 4.23 may partly reflect the LLM rewarding its own idealized version of an activity rather than fidelity to real behavior.
  • Editorial inference: a stronger test of the framework is whether synthetic data generated from one home's sensors improves recognition in a home with a different sensor vocabulary and floorplan; the paper's cross-floorplan result uses a textual representation that partially abstracts away that challenge.
  • Editorial inference: the near-perfect synthetic-to-synthetic accuracy of 98.74 is consistent with the model having learned class-conditional patterns well, but it does not by itself guarantee that those patterns are the ones a clinician would care about.
  • Editorial inference: if the semantic refinement generalizes, the same Insert-Delete-Reorder rule mechanism could be applied to other structured symbolic event streams beyond smart homes, such as clinical workflow logs or industrial sensor alarms.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes ADLGen, a generative framework for symbolic, event-triggered sensor sequences in ambient assisted living. It combines a decoder-only Transformer with sign-based tokenization, symbolic-temporal decoupling, a context- and layout-aware sampling scheme, and an LLM-based generate-evaluate-refine loop. Experiments on the CASAS Aruba dataset report intrinsic gains (MMD2, diversity, validity, semantic quality) and downstream gains (activity recognition, few-shot augmentation). The paper claims that ADLGen closely matches real statistical properties, achieves superior semantic quality, and significantly boosts downstream recognition performance.

Significance. The problem of generating privacy-preserving, realistic ADL sensor data is timely and practically important, and the engineering effort is substantial: code is released, the appendix is detailed, and the evaluation covers both intrinsic and extrinsic dimensions. If the claims were supported by independent and non-circular evaluation, ADLGen would be a useful contribution to the smart-home and human-activity-recognition communities. As presented, however, the central evidence for physical plausibility and semantic quality is undermined by an internal contradiction in the method description and by using the same LLM for generation, refinement, and scoring.

major comments (4)
  1. [Section 3.1, Eq. (6); Appendix H] The method section introduces a floorplan-derived adjacency matrix A_map and rejects or resamples sensor transitions with A_map(|s'_t|, |s_{t-1}|) = 0, stating that this 'ensures physically feasible transitions' and directly supports the Validity Rate metric (Table 3). Appendix H, however, states that the model 'does not explicitly incorporate the spatial topology of the environment' and does not 'directly model the structural information of the floorplan, such as room types ... or spatial adjacency.' These two statements are irreconcilable. If Eq. (6) is implemented as written, Appendix H is false; if it is not, the reported Validity Rate of 1.00 is not produced by the claimed adjacency-constrained mechanism, and the physical-plausibility contribution is unsupported. The authors must either implement and report the adjacency mechanism as described, or remove the corresponding claims, revise contribution 2, and re-interpret the Validity Rate results.
  2. [Section 3.2, Appendix E.4/E.5, Table 3] The headline Semantic Quality metric is computed by Gemini 2.5 Pro (Appendix E.4), and the same LLM generates the refinement rules (Section 3.2.2, Appendix E.5). Table 3 reports a Semantic Quality of 4.67 for refined synthetic sequences versus 4.23 for raw real data, which is interpreted as evidence of superior semantic quality. Because the evaluator and refiner are the same model, the score is not an independent measure of real-world authenticity; it reflects the LLM's internal notion of an 'idealized' activity, as the text itself acknowledges by saying refinement aligns sequences with 'more complete and idealized representations of activities.' This circularity means the semantic-quality claim is not validated. Please provide an external human expert evaluation or an independent rule-based semantic metric, and explicitly discuss the implications of the current self-referential evaluation.
  3. [Appendix A.6, Table 5] The few-shot augmentation experiments filter generated sequences using explicit thresholds: a minimum semantic quality score of 3.5, 100% floorplan validity, and a minimum diversity score of 0.4. Table 5 then reports large macro-F1 improvements (e.g., from 0.76 to 0.99 with 400% synthetic data). Without control conditions—such as adding the same number of real samples, using unfiltered synthetic data, or applying identical filtering to other generators—the improvements cannot be attributed to the generative model's quality rather than to outcome-based selection of easy, already-valid samples. This is especially concerning because the filtering uses the same LLM metric that serves as the paper's primary semantic-quality claim. Please add appropriate baselines and quantify the effect of the filtering step.
  4. [Section 4.1, Table 3] The paper highlights that ADLGen 'surpasses even raw data diversity' (Intra-Set Similarity 0.41 vs. 0.48; Diversity Score 0.86 vs. 0.80). For a generative model whose goal is to match the real distribution, producing sequences that are more diverse than the real data is not self-evidently a positive; it may indicate expansion beyond the true support of the data. The authors should justify why exceeding raw-data diversity is desirable for downstream augmentation, or temper this claim and analyze whether the increased diversity implies a distributional mismatch.
minor comments (5)
  1. [Section 4, Experimental Setup] The text says '5-fold cross-validation (90% training, 10% testing)', which is internally inconsistent: a 5-fold CV protocol normally uses 80% training and 20% testing per fold. Please clarify the actual evaluation protocol.
  2. [Section 3.1, Eq. (6)] The resampling counter notation in Eq. (6) is confusing: 'N=0 if ... RESAMPLE, N+1 otherwise if N ≤ 3' mixes the meaning of N. Please use a separate variable (e.g., n_try) and state the stopping criterion unambiguously.
  3. [Section 3.2 vs. Appendix D.3] The semantic quality decomposition is inconsistent: Section 3.2, Eq. (7) uses phi_F, phi_B, phi_T, while Appendix D.3, Eq. (16) uses alpha_F, alpha_S, alpha_T and refers to phi_F as 'functional' instead of 'fundamental logic'. Please align the notation and definitions.
  4. [Throughout] There are several typos and formatting artifacts, including 'entih er' (Section 3.1), 'Gemimi' (Appendix E.4), 'V ocabulary' (Sections 2 and Appendix C), and 'Vmportant'. A careful proofreading pass is needed.
  5. [Section 1, Contribution 4] The paper claims 'novel metrics' for evaluation, but some are existing measures (e.g., MMD2, entropy-based diversity). Please clarify which components of the evaluation protocol are genuinely new and which are adaptations of existing metrics.

Circularity Check

2 steps flagged · score 6.0 of 10

Validity Rate is a restatement of the model's own adjacency filter, and Semantic Quality is a same-LLM judge/refiner closed loop; downstream recognition results remain independent.

  1. self definitional [Section 3.1 (Eq. 6) and Section 4.1 (Validity Rate definition)]
    "we constrain the generative process to produce only physically feasible transitions—i.e., transitions allowed by the spatial layout of the environment... At each decoding step t, the model proposes a candidate sensor activation s′t. We accept it only if it forms a valid transition with the previous token st−1... If the proposed transition violates physical adjacency... it is rejected and resampled. ... Physical plausibility via Validity Rate↑ using the sensor adjacency matrix from Section 3.1."

    The Validity Rate is computed with the same adjacency matrix Amap that Eq. 6 uses as a hard rejection filter. If the adjacency constraint is enforced during sampling, every emitted sequence satisfies Amap(...)=1 by construction, so the reported Validity Rate of 1.00 is not an independent measurement of physical plausibility—it simply restates the generation constraint. If the constraint is not actually enforced (as Appendix H later says the model 'does not directly model ... spatial adjacency'), then the perfect rate is vacuous. Either way, Table 3's 'physical plausibility' evidence does not independently establish the claimed property; the metric and the mechanism share the same definition.

  2. self definitional [Section 3.2 / Appendix E.4 and E.5 / Table 3]
    "In our study, the LLM from Google Gemini family: Gemimi 2.5 Pro Preview ... You are an AI assistant specialized in evaluating the semantic quality of sensor event sequences ... You are an AI assistant tasked with refining sensor event sequences representing human Activities of Daily Living (ADLs)... You now must generate and apply executable rules to correct semantic violations... refine all sequences in the dataset that received scores below 5."

    The Semantic Quality score is assigned by the same LLM (Gemini 2.5 Pro Preview) that the refinement prompt instructs to generate and apply correction rules targeting the exact issues the scoring rubric penalizes. The refinement step therefore optimizes the judge's own scoring criteria, and the same judge then reports the post-optimization score (3.79 to 4.67, exceeding raw real data at 4.23). This is not independent evidence of behavioral authenticity; it demonstrates that the LLM can make sequences conform to its own stated preferences. The paper itself attributes the gain to 'more complete and idealized representations of activities,' confirming the metric rewards the evaluator's ideal rather than fidelity to real data.

full rationale

ADLGen's strongest independent evidence is the extrinsic, task-based evaluation: TSTR training on synthetic data reaches MacroF1 72.90 versus 79.73 for real-trained models (Table 4), and few-shot augmentation lifts the 15-sample MacroF1 from 0.76 to 0.99 (Table 5). These results are not forced by construction and are the main reason the paper's overall contribution is not entirely circular. MMD2 and diversity are also computed directly between generated and real sequences and are not self-referential. The circularity is concentrated in two of the paper's four proposed evaluation dimensions. First, the Validity Rate in Table 3 is defined using the same adjacency matrix Amap that Eq. 6 uses to reject invalid transitions; if the sampler works, 1.00 is guaranteed by construction, and if it does not, the number is vacuous. Appendix H compounds this by stating the model 'does not directly model ... spatial adjacency,' contradicting Eq. 6; that is a correctness/integrity problem rather than a circularity per se, but it removes the independent support the physical-plausibility claim would need. Second, the Semantic Quality metric is a closed loop: Gemini 2.5 Pro both scores the sequences (Appendix E.4) and, via the refinement prompt, rewrites them to fix the very issues it scored (Appendix E.5). Reporting 4.67 after refinement, above raw real data's 4.23, measures self-consistency with the LLM's own rubric rather than fidelity to real behavior. These two self-referential metrics support the 'physically plausible' and 'semantically rich' claims, so the paper earns a 6 rather than a lower score, but the downstream recognition benchmark prevents a higher score because the main practical claim is externally tested.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The most consequential free parameters are the quality filter thresholds in Appendix A.6, which directly select the data used in the headline augmentation results. The axioms include the reliance on CASAS annotations, the adjacency matrix as a complete model of physical movement, and the LLM as a valid semantic oracle. No new physical entities are introduced.

free parameters (6)
  • base temperature τ_base = 0.7
    Chosen by hand for nucleus sampling; controls exploration/exploitation balance in generation (Appendix A.4, Eq. 5).
  • repetition penalty rp/gamma = 1.1
    Tunable hyperparameter penalizing overused sensors; set in Appendix A.4.
  • length-aware temperature scale (τ_min, τ_max, L_min=3, L_max=100) = not fully specified
    Defined in Eq. (5) with τ_min and τ_max not given numeric values; affects diversity-vs-coherence tradeoff.
  • diversity factor beta = in {1, 1.2}
    Increases temperature when sensor diversity is low (Eq. 5).
  • adjacency resampling attempts N = 3
    Number of reject/resample attempts for physically invalid transitions (Eq. 6).
  • quality filter thresholds for augmentation = semantic >= 3.5, validity = 100%, diversity >= 0.4
    Appendix A.6: generated sequences are filtered on the same metrics used to report success, directly affecting Tables 4 and 5.
assumptions (5)
  • domain assumption CASAS Aruba activity annotations are correct ground truth and excluding 'Other Activity' and temperature sensors does not bias the distribution.
    Used throughout preprocessing (Appendix A.2, Algorithm 1).
  • domain assumption The floorplan adjacency matrix Amap fully represents physically feasible transitions, and rejecting non-adjacent transitions does not discard legitimate behaviors such as moving through unmonitored areas.
    Invoked in Section 3.1 Eq. (6) and in the validity-rate metric.
  • ad hoc to paper The LLM's semantic quality score is a valid measure of real-world ADL authenticity, and the same LLM's refinement rules improve rather than distort sequences.
    Core to Section 3.2 and Appendix E; the paper's own admission that refined data align with more complete and idealized representations (Section 4.1) indicates the metric may reward idealized behavior.
  • domain assumption A sliding window of 10 events and max sequence length 100 preserve enough context for activity modeling.
    Appendix A.2 and Section 3.1 define the input/output format.
  • standard math The MMD estimator with an unspecified kernel and sequence embedding is a valid measure of distributional fidelity.
    Section 4.1 and Appendix D.2 use MMD; the kernel and how variable-length sequences are embedded are not detailed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ADLGen: Synthesizing Symbolic, Event-Triggered Sensor Sequences for Human Activity Modeling." pith.science (2026). https://pith.science/paper/FNUN6EXB

@misc{pith2026250517987,
  author       = {Pith},
  title        = {Pith review of: ADLGen: Synthesizing Symbolic, Event-Triggered Sensor Sequences for Human Activity Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FNUN6EXB}},
  note         = {Machine review of arXiv:2505.17987}
}
read the original abstract

Real world collection of Activities of Daily Living data is challenging due to privacy concerns, costly deployment and labeling, and the inherent sparsity and imbalance of human behavior. We present ADLGen, a generative framework specifically designed to synthesize realistic, event triggered, and symbolic sensor sequences for ambient assistive environments. ADLGen integrates a decoder only Transformer with sign based symbolic temporal encoding, and a context and layout aware sampling mechanism to guide generation toward semantically rich and physically plausible sensor event sequences. To enhance semantic fidelity and correct structural inconsistencies, we further incorporate a large language model into an automatic generate evaluate refine loop, which verifies logical, behavioral, and temporal coherence and generates correction rules without manual intervention or environment specific tuning. Through comprehensive experiments with novel evaluation metrics, ADLGen is shown to outperform baseline generators in statistical fidelity, semantic richness, and downstream activity recognition, offering a scalable and privacy-preserving solution for ADL data synthesis.

Figures

Figures reproduced from arXiv: 2505.17987 by the authors.

Figure 1
Figure 1. Semantic Quality evaluation (1-5 scale) across different daily activities. Despite the clinical and commercial importance of ADL modeling, large-scale data collection remains constraint by three intertwined barriers. First, continuous monitoring in private spaces raises serious privacy concerns [23, 24]. Sec￾ond, installing and maintaining dense sensor net￾works is time-consuming and costly and labor￾intensive, espe… view at source ↗
Figure 2
Figure 2. Our Two-Stage Synthetic Activity Data Generation Framework. (a) Pretrain Stage: Transformer learns ADL patterns from sign-based, decoupled encoded sensor data. (b) Inference Stage: Transformer generates sequences which are subsequently refined semantically by a large language model to enhance coherence and contextual relevance. records the exact timestamp of the event (see [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. LLM evaluation framework with refinement pipeline. Two-Tiered Evaluation. The evaluation is organized in a two-tiered pipeline (Fig￾ure 3), inspired by Multi￾Tiered Systems of Support (MTSS) [57]. In Test A, the LLM validates fundamental logic by checking physical and logical consistency. Only sequences that pass Test A proceed to Test B, where the LLM jointly evaluates behav￾ioral coherence and temporal consistency… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: t-SNE visualizations of Bi-LSTM classifier (trained [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Sensor layout of Aruba floorplan Activity Annotation Process. Activities in the dataset were annotated using a combination of resident self-reporting and post-hoc verification by researchers. The annotation process involved marking the start and end times of specific a…
Figure 6
Figure 6. Figure 6: Distribution of sensor event intervals. 1. Information Asymmetry: Sensor (ID+status) patterns typically carry the most discriminative information for ADL identification, while temporal features often remain static throughout an activity (e.g., an entire "bed to toilet"…
Figure 7
Figure 7. Figure 7: Confusion matrices comparing cross-domain generalization. (a) Shows the classifier trained [PITH_FULL_IMAGE:figures/full_fig_p024_7.png]
Figure 8
Figure 8. Figure 8: t-SNE visualizations of data embeddings from the Bi-LSTM classifier. Both plots co [PITH_FULL_IMAGE:figures/full_fig_p025_8.png]
Figure 9
Figure 9. Figure 9: t-SNE visualizations of data embeddings from the Bi-LSTM classifier. Both plots co [PITH_FULL_IMAGE:figures/full_fig_p025_9.png]
Figure 11
Figure 11. Figure 11: Aruba Floorplan common standardized activity taxonomy, restricting our evaluation exclusively to the intersection of activity classes present in both environments. Statistical rigor was preserved through implementation of standard cross-validation procedures with cons…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

66 extracted references · 60 canonical work pages

  1. [1]

    M. H. Arshad, M. Bilal, and A. Gani. Human activity recognition: Review, taxonomy and open challenges. Sensors, 22(17):6463, 2022

    work page 2022

  2. [2]

    X. Ye, K. Sakurai, N. C. Nair, and K. I. Wang. Machine learning techniques for sensor-based human activity recognition with data heterogeneity-a review. Sensors, 24(24):7975, 2024

    work page 2024

  3. [3]

    Ambient assisted living: a review of technologies, methodologies and future perspectives for healthy aging of population

    Grazia Cicirelli, Roberto Marani, Antonio Petitti, Annalisa Milella, and Tiziana D’orazio. Ambient assisted living: a review of technologies, methodologies and future perspectives for healthy aging of population. Sensors, 21(10):3549, 2021

    work page 2021

  4. [4]

    Lee, and Soon Ju Kang

    Ho-Min Kang, C. Lee, and Soon Ju Kang. A smart device for non-invasive adl estimation through multi-environmental sensor fusion. Scientific Reports, 13, 2023

    work page 2023

  5. [5]

    Gathering activities of daily living data for elderly care in network deficient environments.IEEE Access, 12:121144–121155, 2024

    Huan-Bang Li, Lin Shan, Takeshi Matsumura, and Yasushi Fuwa. Gathering activities of daily living data for elderly care in network deficient environments.IEEE Access, 12:121144–121155, 2024

    work page 2024

  6. [6]

    A systematic review of wearable sensors and iot-based monitoring applications for older adults–a focus on ageing population and independent living

    Mirza Mansoor Baig, Shereen Afifi, Hamid GholamHosseini, and Farhaan Mirza. A systematic review of wearable sensors and iot-based monitoring applications for older adults–a focus on ageing population and independent living. Journal of medical systems, 43:1–11, 2019

    work page 2019

  7. [7]

    A multisen- sor data-fusion approach for adl and fall classification

    Bruno Andò, Salvatore Baglio, Cristian Orazio Lombardo, and Vincenzo Marletta. A multisen- sor data-fusion approach for adl and fall classification. IEEE Transactions on Instrumentation and Measurement, 65(9):1960–1967, 2016

    work page 1960

  8. [8]

    Fine-grained activity of daily living (adl) recognition through heterogeneous sensing systems with com- plementary spatiotemporal characteristics

    Shijia Pan, Mario Berges, Juleen Rodakowski, Pei Zhang, and Hae Young Noh. Fine-grained activity of daily living (adl) recognition through heterogeneous sensing systems with com- plementary spatiotemporal characteristics. Frontiers in Built Environment, V olume 6 - 2020, 2020

    work page 2020

Show all 66 references
  1. [9]

    Monitoring activities of daily living of the elderly and the potential for its use in telecare and telehealth: a review

    Hulya Gokalp and Malcolm Clarke. Monitoring activities of daily living of the elderly and the potential for its use in telecare and telehealth: a review. TELEMEDICINE and e-HEALTH, 19(12):910–923, 2013

    2013

  2. [10]

    A sequence-to-sequence model-based deep learning approach for recognizing activity of daily living for senior care

    Hongyi Zhu, Hsinchun Chen, and Randall Brown. A sequence-to-sequence model-based deep learning approach for recognizing activity of daily living for senior care. Journal of biomedical informatics, 84:148–158, 2018

    2018

  3. [11]

    B. Andò, S. Baglio, C. O. Lombardo, and V . Marletta. A multisensor data-fusion approach for adl and fall classification. IEEE Transactions on Instrumentation and Measurement, 65:1960–1967, 2016

    1960

  4. [12]

    Vermeulen, Jacques C L Neyens, E

    J. Vermeulen, Jacques C L Neyens, E. van Rossum, M. Spreeuwenberg, and Luc de Witte. Predicting adl disability in community-dwelling elderly people using physical frailty indicators: a systematic review. BMC Geriatrics, 11:33 – 33, 2011

    2011

  5. [13]

    Luperto, J

    M. Luperto, J. Monroy, Jennifer Renoux, Francesca Lunardini, Nicola Basilico, M. Bulgheroni, A. Cangelosi, Matteo Cesari, Manuel Cid, A. Ianes, J. Gonzalez-Jimenez, A. Kounoudes, David Marí, Victor Prisacariu, A. Savanovic, S. Ferrante, and N. A. Borghese. Integrating social a...

    2022

  6. [14]

    Marija Stojchevska, Mathias De Brouwer, Martijn Courteaux, Bram Steenwinckel, Sofie Van Hoecke, and F. Ongenae. Unlocking the potential of smartphone and ambient sensors for adl detection. Scientific Reports, 14, 2024

    2024

  7. [15]

    ElHady, Stephan M

    Nancy E. ElHady, Stephan M. Jonas, Julien Provost, and V . Senner. Sensor failure detection in ambient assisted living using association rule mining. Sensors (Basel, Switzerland), 20, 2020

    2020

  8. [16]

    Casas: A smart home in a box

    Diane J Cook, Aaron S Crandall, Brian L Thomas, and Narayanan C Krishnan. Casas: A smart home in a box. Computer, 46(7):62–69, 2012. 31

    2012

  9. [17]

    Aras human activity datasets in multiple homes with multiple residents

    Hande Alemdar, Halil Ertan, Ozlem Durmaz Incel, and Cem Ersoy. Aras human activity datasets in multiple homes with multiple residents. In 2013 7th International Conference on Pervasive Computing Technologies for Healthcare and Workshops, pages 232–235. IEEE, 2013

    2013

  10. [19]

    Accurate action recommen- dation for smart home via two-level encoders and commonsense knowledge

    Hyunsik Jeon, Jongjin Kim, Hoyoung Yoon, Jaeri Lee, and U Kang. Accurate action recommen- dation for smart home via two-level encoders and commonsense knowledge. In Mohammad Al Hasan and Li Xiong, editors, Proceedings of the 31st ACM International Conference on Infor- mation &...

    2022

  11. [20]

    Youhome system and dataset: Making your home know you better

    Junhao Pan, Zehua Yuan, Xiaofan Zhang, and Deming Chen. Youhome system and dataset: Making your home know you better. IEEE International Symposium on Smart Electronic Systems (IEEE - iSES), 2022

    2022

  12. [21]

    Ascc activities of daily living dataset, 2023

    Fei Liang. Ascc activities of daily living dataset, 2023

    2023

  13. [22]

    Activity recognition using semi- markov models on real world smart home datasets

    TLM Van Kasteren, Gwenn Englebienne, and Ben JA Kröse. Activity recognition using semi- markov models on real world smart home datasets. Journal of ambient intelligence and smart environments, 2(3):311–325, 2010

    2010

  14. [23]

    J. M. Mendes, A. Barbar, and M. Refaie. Synthetic data generation: a privacy-preserving approach to accelerate rare disease research. Frontiers in Digital Health, 7, 2025

    2025

  15. [24]

    Living alone among older adults in canada and the us

    Sharon M Lee and Barry Edmonston. Living alone among older adults in canada and the us. In Healthcare, volume 7, page 68. MDPI, 2019

    2019

  16. [25]

    H. Chen, C. Gouin-Vallerand, K. Bouchard, S. Gaboury, M. Couture, N. Bier, and S. Giroux. Enhancing human activity recognition in smart homes with self-supervised learning and self- attention. IEEE Sensors Journal, 2024. Advance online publication

    2024

  17. [26]

    D’souza, M

    A. D’souza, M. Swetha, and S. Sarawagi. Synthetic tabular data generation for imbalanced classification: The surprising effectiveness of an overlap class. In The Thirty-Ninth AAAI Conference on Artificial Intelligence (AAAI-25), 2025

    2025

  18. [27]

    Smote for time series classification: A case study in human activity recognition

    Alberto Fernández, Carlos José Carmona, María José del Jesus, and Francisco Herrera. Smote for time series classification: A case study in human activity recognition. Neurocomputing, 275:1513–1525, 2018

    2018

  19. [28]

    Multivariate time series data imputation using attention-based mechanism

    Jingqi Zhao, Chuitian Rong, Chunbin Lin, and Xin Dang. Multivariate time series data imputation using attention-based mechanism. Neurocomputing, 542:126238, 2023

    2023

  20. [29]

    Rndiff: Rainfall nowcasting with condition diffusion model

    Xudong Ling, Chaorong Li, Fengqing Qin, Peng Yang, and Yuanyuan Huang. Rndiff: Rainfall nowcasting with condition diffusion model. Pattern Recognit., 160:111193, 2025

    2025

  21. [30]

    Fin-gan: forecasting and classifying financial time series via generative adversarial networks

    Milena Vuleti´c, Felix Prenzel, and Mihai Cucuringu. Fin-gan: forecasting and classifying financial time series via generative adversarial networks. Quantitative Finance, 24:175 – 199, 2024

    2024

  22. [31]

    Eldersim: A synthetic data generation platform for human action recognition in eldercare applications

    Hochul Hwang, Cheongjae Jang, Geonwoo Park, Junghyun Cho, and Ig-Jae Kim. Eldersim: A synthetic data generation platform for human action recognition in eldercare applications. IEEE Access, 11:9279–9294, 2021

    2021

  23. [32]

    Activity scenarios simulation by discovering knowledge through activities of daily living datasets

    Swe Nwe Nwe Htun, Shusaku Egami, and Koji Fukuda. Activity scenarios simulation by discovering knowledge through activities of daily living datasets. SICE Journal of Control, Measurement, and System Integration, 17(1):87–105, 2024

    2024

  24. [33]

    Transformer Hawkes process

    Simiao Zuo, Haoming Jiang, Zichong Li, Tuo Zhao, and Hongyuan Zha. Transformer Hawkes process. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 11692–1...

    2020

  25. [34]

    Self-attentive hawkes process

    Qiang Zhang, Aldo Lipani, Omer Kirnap, and Emine Yilmaz. Self-attentive hawkes process. In International conference on machine learning, pages 11183–11193. PMLR, 2020

    2020

  26. [35]

    Time-series generative adversarial networks

    Jinsung Yoon, Daniel Jarrett, and Mihaela Van der Schaar. Time-series generative adversarial networks. Advances in neural information processing systems, 32, 2019

    2019

  27. [36]

    Cot-gan: Generating sequential data via causal optimal transport

    Tianlin Xu, Li Kevin Wenliang, Michael Munn, and Beatrice Acciaio. Cot-gan: Generating sequential data via causal optimal transport. Advances in neural information processing systems, 33:8798–8809, 2020

    2020

  28. [37]

    Predict, refine, synthesize: Self-guiding diffusion models for probabilistic time series forecasting

    Marcel Kollovieh, Abdul Fatir Ansari, Michael Bohlke-Schneider, Jasper Zschiegner, Hao Wang, and Yuyang Bernie Wang. Predict, refine, synthesize: Self-guiding diffusion models for probabilistic time series forecasting. Advances in Neural Information Processing Systems, 36:2834...

    2023

  29. [38]

    Timeldm: La- tent diffusion model for unconditional time series generation

    Jian Qian, Bingyu Xie, Biao Wan, Minhao Li, Miao Sun, and Patrick Yin Chiang. Timeldm: La- tent diffusion model for unconditional time series generation. arXiv preprint arXiv:2407.04211, 2024

    2024 arXiv

  30. [39]

    Layout-agnostic human activity recognition in smart homes through textual descriptions of sensor triggers (tdost)

    Megha Thukral, Sourish Gunesh Dhekane, Shruthi K Hiremath, Harish Haresamudram, and Thomas Ploetz. Layout-agnostic human activity recognition in smart homes through textual descriptions of sensor triggers (tdost). Proceedings of the ACM on Interactive, Mobile, Wearable and Ubi...

    2025

  31. [40]

    A sequential deep learning application for recognising human activities in smart homes

    Daniele Liciotti, Michele Bernardini, Luca Romeo, and Emanuele Frontoni. A sequential deep learning application for recognising human activities in smart homes. Neurocomputing, 2019

    2019

  32. [41]

    Fine-grained activity of daily living (adl) recognition through heterogeneous sensing systems with comple- mentary spatiotemporal characteristics

    Shijia Pan, Mario Berges, Juleen Rodakowski, Pei Zhang, and Hae Young Noh. Fine-grained activity of daily living (adl) recognition through heterogeneous sensing systems with comple- mentary spatiotemporal characteristics. Frontiers in Built Environment, 6:560497, 2020

    2020

  33. [42]

    Multimodal generation of novel action appearances for synthetic-to-real recognition of activities of daily living

    Zhivko Marinov, David Schneider, Alina Roitberg, and Rainer Stiefelhagen. Multimodal generation of novel action appearances for synthetic-to-real recognition of activities of daily living. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), page...

    2022

  34. [43]

    A quantitative analysis of activities of daily living: Insights into improving functional independence with assistive robotics

    Laura Petrich, Jun Jin, Masood Dehghan, and Martin Jagersand. A quantitative analysis of activities of daily living: Insights into improving functional independence with assistive robotics. In 2022 International Conference on Robotics and Automation (ICRA), pages 6999–7006. IEEE, 2022

    2022

  35. [44]

    Smote: Synthetic minority over-sampling technique

    Nitesh V Chawla, Kevin W Bowyer, Lawrence O Hall, and W Philip Kegelmeyer. Smote: Synthetic minority over-sampling technique. Journal of Artificial Intelligence Research, 16:321– 357, 2002

    2002

  36. [45]

    Constrained diffusion models via dual training

    Shervin Khalafi, Dongsheng Ding, and Alejandro Ribeiro. Constrained diffusion models via dual training. arXiv preprint arXiv:2408.15094, 2024

    2024 arXiv

  37. [46]

    Stavropoulos, S

    Giorgos Giannios, Lampros Mpaltadoros, Vasilis Alepopoulos, Margarita Grammatikopoulou, Thanos G. Stavropoulos, S. Nikolopoulos, Ioulietta Lazarou, Magdalini Tsolaki, and Y . Kompat- siaris. A semantic framework to detect problems in activities of daily living monitored throug...

    2024

  38. [47]

    Chatterjee, Martin W

    A. Chatterjee, Martin W. Gerdes, Andreas Prinz, Michael A Riegler, and Santiago G Martinez. Semantic representation and comparative analysis of physical activity sensor observations using mox2-5 sensor in real and synthetic datasets: a proof-of-concept-study. Scientific Report...

    2024

  39. [48]

    Cgxplain: Rule-based deep neural network explanations using dual linear programs

    Konstantin Hemker, Zohreh Shams, and Mateja Jamnik. Cgxplain: Rule-based deep neural network explanations using dual linear programs. In Proceedings of the International Workshop on Trustworthy Machine Learning for Healthcare (TML4H) , pages 60–72. Springer Nature Switzerland,...

    2023

  40. [49]

    Star-transformer: A spatio-temporal cross attention transformer for human action recognition

    Dongkwan Ahn, Seungjun Kim, Hyeonwoo Hong, and ByoungChul Ko. Star-transformer: A spatio-temporal cross attention transformer for human action recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3330–3339. IEEE, 2023

    2023

  41. [50]

    Auggpt: Leveraging chatgpt for text data augmentation

    Haixing Dai, Zhengliang Liu, Wenxiong Liao, Xiaoke Huang, Yihan Cao, Zihao Wu, Lin Zhao, Shaochen Xu, Fang Zeng, Wei Liu, et al. Auggpt: Leveraging chatgpt for text data augmentation. IEEE Transactions on Big Data, 2025

    2025

  42. [51]

    A systematic assessment of openai o1-preview for higher order thinking in education

    Ehsan Latif, Yifan Zhou, Shuchen Guo, Yizhu Gao, Lehong Shi, Matthew Nayaaba, Gyeonggeon Lee, Liang Zhang, Arne Bewersdorff, Luyang Fang, et al. A systematic assessment of openai o1-preview for higher order thinking in education. arXiv preprint arXiv:2410.21287, 2024

    arXiv

  43. [52]

    Large language model-driven structured output: A comprehensive benchmark and spatial data generation framework

    Diya Li, Yue Zhao, Zhifang Wang, Calvin Jung, and Zhe Zhang. Large language model-driven structured output: A comprehensive benchmark and spatial data generation framework. ISPRS International Journal of Geo-Information, 2024

    2024

  44. [53]

    Are llms good at structured outputs? a benchmark for evaluating structured output capabilities in llms

    Yu Liu, Duantengchuan Li, Kaili Wang, Zhuoran Xiong, Fobo Shi, Jian Wang, Bing Li, and Bo Hang. Are llms good at structured outputs? a benchmark for evaluating structured output capabilities in llms. Inf. Process. Manag., 61:103809, 2024

    2024

  45. [54]

    Matez-Bandera, J

    Jesus Moncada-Ramirez, J. Matez-Bandera, J. Gonzalez-Jimenez, and J. Ruiz-Sarmiento. Agentic workflows for improving large language model reasoning in robotic object-centered planning. Robotics, 2025

    2025

  46. [55]

    E, and Meina Song

    Zecheng Zhan, H. E, and Meina Song. Leveraging large language model for enhanced text-to-sql parsing. IEEE Access, 13:30497–30504, 2025

    2025

  47. [56]

    A kernel method for the two-sample-problem

    Arthur Gretton, Karsten Borgwardt, Malte Rasch, Bernhard Schölkopf, and Alex Smola. A kernel method for the two-sample-problem. Advances in neural information processing systems, 19, 2006

    2006

  48. [57]

    Multi-tiered systems of support and evidence-based practices

    Karen C Stoiber and Maribeth Gettinger. Multi-tiered systems of support and evidence-based practices. In Handbook of response to intervention: The science and practice of multi-tiered systems of support, pages 121–141. Springer, 2015

    2015

  49. [58]

    Ortigosa

    Marcos Lupión, Federico Cruciani, Ian Cleland, Chris Nugent, and Pilar M. Ortigosa. Data augmentation for human activity recognition with generative adversarial networks. IEEE Journal of Biomedical and Health Informatics, 28(4):2350–2361, 2024

    2024

  50. [59]

    Generative resident separation and multi-label classification for multi-person activity recognition

    Xi Chen, Julien Cumin, Fano Ramparany, and Dominique Vaufreydaz. Generative resident separation and multi-label classification for multi-person activity recognition. In 2024 IEEE International Conference on Pervasive Computing and Communications Workshops and other Affiliated ...

    2024

  51. [60]

    Attention- based bidirectional long short-term memory networks for relation classification

    Peng Zhou, Wei Shi, Jun Tian, Zhenyu Qi, Bingchen Li, Hongwei Hao, and Bo Xu. Attention- based bidirectional long short-term memory networks for relation classification. In Proceedings of the 54th annual meeting of the association for computational linguistics (volume 2: Short...

    2016

  52. [61]

    Real-valued (medical) time series generation with recurrent conditional gans

    Cristóbal Esteban, Stephanie L Hyland, and Gunnar Rätsch. Real-valued (medical) time series generation with recurrent conditional gans. arXiv preprint arXiv:1706.02633, 2017

    2017 arXiv

  53. [62]

    Morris Chang

    Mohammad Al-Rubaie and J. Morris Chang. Privacy-preserving machine learning: Threats and solutions. IEEE Security & Privacy, 17(2):49–58, 2019

    2019

  54. [63]

    X. Li, J. Luo, and R. Younes. Activitygan: Generative adversarial networks for data augmen- tation in sensor-based human activity recognition. In Adjunct Proceedings of the 2020 ACM International Joint Conference on Pervasive and Ubiquitous Computing & Proceedings of the 2020 ...

    2020

  55. [64]

    Unsupervised domain adaptation by backpropagation

    Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In Proceedings of the 32nd International Conference on Machine Learning (ICML 2015) , volume 37 of Proceedings of Machine Learning Research, pages 1180–1189. PMLR, 2015

    2015

  56. [65]

    Measuring the quality of synthetic data for use in competitions

    James Jordon, Jinsung Yoon, and Mihaela van der Schaar. Measuring the quality of synthetic data for use in competitions. arXiv preprint arXiv:1806.11345, 2018

    2018 arXiv

  57. [66]

    Gender shades: Intersectional accuracy disparities in commercial gender classification

    Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Proceedings of the Conference on Fairness, Accountability and Transparency (FAT* 2018), volume 81 of Proceedings of Machine Learning Research, pages 77–9...

    2018

  58. [67]

    The global landscape of AI ethics guidelines

    Anna Jobin, Marcello Ienca, and Effy Vayena. The global landscape of AI ethics guidelines. Nature Machine Intelligence, 1(9):389–399, 2019. 35

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.