acceptodds
Under review as a conference paper at ICLR 2027

Cluster-LLM: Clustering-Guided LLM-Based Data Augmentation for Anomaly Detection in Smart Home

Abstract

Anomaly detection in activities of daily living (ADL) within smart home environments is critical for supporting the safety and well-being of older adults living independently. However, the scarcity of annotated anomaly data and the inherent behavioral heterogeneity among residents pose significant challenges to developing robust detection models. To address these issues, we propose a novel clustering-guided LLM-based data augmentation framework (Cluster-LLM) for ADL anomaly detection. This framework first leverages clustering algorithms to mine differential behavioral patterns from raw sensor data, and then generates lifelike synthetic activity sequences matching the characteristics of each behavioral group through LLM reasoning. We conduct extensive evaluations on four clustering approaches, namely K-means, K-Medoids, DBSCAN and deep embedded clustering (DEC), using two standard smart home datasets. Extensive experiments with four one-class classifiers and seven evaluation metrics demonstrate that Cluster-LLM achieves satisfactory performance in anomaly detection in smart home. Specifically, Cluster-LLM boosts the best F1-score on the Kyoto7 dataset from 0.444 to 0.667, representing a 50.2% relative improvement over baselines. On the unsupervised benchmark, the proposed framework enhances accuracy across diverse clustering backbones, while maintaining robust F1 gains within a moderate augmentation scale.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM, Data Augmentation, Anomaly Detection, Clustering, Smart Home

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_79NLOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Anomaly detection over activities of daily living in smart homes suffers from scarce anomaly labels and heterogeneous resident behavior. The paper proposes Cluster-LLM, which clusters sensor-sequence features (K-means, K-Medoids, DBSCAN, or DEC), summarizes each cluster as a statistical profile, and prompts an LLM (Qwen-plus) with that profile, the sensor topology, activity specifications, and anomaly definitions to generate synthetic normal and anomalous sensor-event sequences. The synthetic data is added to the real training set of four one-class detectors (OCSVM, iForest, LOF, Elliptic Envelope). Validation compares each detector with and without augmentation on CASAS Kyoto7 (scripted activities with injected errors) and HH111 (61 days, day-level labels derived by the authors' rules), reporting a phone-call F1 increase from 0.444 to 0.667 on Kyoto7 and volume-dependent, mostly mixed effects on HH111.

Critical strengths
  • The problem is practically relevant, since anomaly labels in ambient-sensor ADL data are scarce, and the code and data are released.
Critical weaknesses
  • The effect of cluster conditioning is not isolated. The only comparison is against no augmentation; a baseline using the same LLM pipeline without cluster profiles, and standard oversampling (SMOTE/ADASYN) on the same feature vectors, are missing, so the gains cannot be attributed to clustering.
  • The evaluation does not establish a reliable gain. On Kyoto7, the phone-call test set has 8 instances (3 anomalous) and results are reported without variance. On HH111, the best augmented F1 (0.274, iForest at D = 50, Table D1) is close to the best unaugmented F1 (0.267, OCSVM).
Questions for authors
  1. How does Cluster-LLM compare with the same pipeline without cluster profiles, and with SMOTE/ADASYN on the same features, at matched synthetic volume?
  2. Table 1 reports single values although stratified cross-validation is used. What are the mean and variance across folds and repeated generation runs?
Reviewer_WdkyOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper explores clustering-based sensor data augmentation strategies for activities of daily living (ADL) with the power of large language models, including Qwen. Since anomaly detection in the context of ADL requires the model to be discriminative with scarce labeled samples, previous approaches have adopted a pre-trained LLM to generate synthetic training samples for normal and abnormal activities. These approaches nonetheless fail to capture fine-grained statistics of residents’ activities, leading to poor generalization to unseen normal and abnormal samples. To overcome this challenge, this work introduces a novel clustering-oriented training data augmentation that divides the original training set into several homogeneous subsets and then prompts the statistics of each cluster into a pre-trained LLM to generate realistic and diverse sensor signals. Experimental results on Kyoto7 and HH111 show that the proposed method mostly performs well compared with non-augmented baselines.

Critical strengths

S1: Presentation. This paper is well-written and easy to follow. Specifically, the introduction and methodology sections present a simple narrative for this work, and the motivation is clear. Some diagrams (e.g., Fig. 1, Fig. A1) provide a concise overview of the paper’s concept for readers unfamiliar with this field.

S2: Effectiveness compared with baselines. Experimental results in Tab. 1 and Tab. 2 demonstrate that the proposed method yields substantial accuracy gains compared with the “no-synthetic data” baselines. While such effectiveness differs across datasets and clustering methodologies, LLM-based generation with clustering seems to contribute to the accuracy improvements in anomaly detection for ADL.

Critical weaknesses

W1: Comparison with other LLM-based generation methodologies. This paper lacks a comparison with other data augmentation approaches in the field of anomaly detection for ADL. Specifically, several previous attempts, including SMOTE [1], ADASYN [2], and AgentSense [3], have been introduced in the introduction section (see L45-L50). Comparing these methodologies will provide a clearer understanding of the proposed method’s effectiveness. It should be noted that [3] has also evaluated its method on Kyoto7. While evaluation or training protocols may differ, it is recommended to compare these methods under fair conditions.

W2: Evaluation with other anomaly detection approaches. Conceptually, we can adopt more recent anomaly detection approaches with the proposed training data augmentation framework. For example, reconstruction-based [4, 5] perform anomaly detection by simply training an autoencoder on normal training data (i.e., unsupervised setting).

W3: In-depth analysis. This paper shows that conditioning using per-cluster stats improves detection accuracy compared with non-augmented baselines. However, several important questions are still not resolved: (a) how does the accuracy drop when we prompt an LLM without cluster information?, (b) how does the accuracy vary across different synthetic anomalous ratios?, (c) do the extracted features from real samples truly exhibit multi-modal properties where non-linear clustering would perform better?, and (d) are there any differences between generating normal and anomalous samples with the proposed cluster-informed prompting?

Questions for authors

Q1: How about the accuracy of other competitors? It is recommended to report the accuracy of other augmentation-based approaches, including SMOTE, ADASYN, and AgentSense.

Q2: Can the proposed method be combined with other anomaly detection approaches? It is suggested to incorporate the reconstruction-based framework for estimating anomalies in an unsupervised setting (r=0).

Please also see the weakness section, especially for W3.

Reviewer_exvgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Cluster-LLM, a data augmentation framework for smart-home activity anomaly detection. It clusters real behavioral data and uses cluster-specific profiles, sensor topology, and activity constraints to guide Qwen-plus in generating synthetic sensor sequences. The generated data are combined with real training data to train anomaly detectors. Experiments cover Kyoto7 and HH111, four clustering methods, four detectors, and several generation volumes. The reported improvements vary across settings, and excessive augmentation can reduce performance.

Critical strengths
  • The paper addresses a practical problem: limited anomaly data and heterogeneous daily routines. Conditioning generation on observed behavioral groups offers a reasonable way to adapt synthetic data to the target environment.

  • The experiments cover multiple clustering methods, detectors, and augmentation volumes. Reporting cases where augmentation hurts performance helps readers understand the method’s limitations.

Critical weaknesses
  • The contribution of clustering is not isolated. The experiments compare against no augmentation, but not against LLM generation using a global behavioral profile under the same budget. The reported gains could therefore come from additional synthetic data rather than cluster conditioning. This missing comparison makes it difficult to assess the paper’s main contribution.

  • The evaluation does not establish generalization beyond the data source used for augmentation, and possible leakage needs clarification. Kyoto7 shares participants between training and evaluation. For HH111, the supplied preprocessing and clustering scripts appear to process the full real stream, although the paper specifies separate training and test days. If the test days influenced generation, the evaluation would no longer be independent. A test on residents or homes excluded from the augmentation process would provide stronger evidence of generalization.

  • The quality of the generated data is not directly evaluated. The paper claims that cluster conditioning preserves behavioral diversity and reduces distribution shift, but provides no quantitative comparison of real and synthetic distributions or their behavioral coverage. Improved detection scores alone do not establish these claimed properties.

Questions for authors
  • With the same generation budget and prompt components, how does cluster-conditioned generation compare with generation using a single global behavioral profile?

  • Were the evaluation data excluded before all data-dependent preprocessing, clustering, profile construction, and generation? Please clarify this specifically for the 51 HH111 test days. Can the method also improve detection on residents or homes that contributed no data to generation?

  • What were the feature-space consistency checks mentioned in the paper? Please report their results and compare the fidelity and diversity of cluster-conditioned generation against a global-prompt baseline, evaluating normal and anomalous sequences separately.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.