Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Exploring the impact of Optimised Hyperparameters on Bi-LSTM-based Contextual Anomaly Detector

Aafan Ahmad Toor, Jia-Chun Lin, Ernst Gunnar Gran

TL;DR

This work tackles contextual anomaly detection in IoT smart-home time-series by extending UoCAD with offline Bi-LSTM hyperparameter optimisation (UoCAD-OH). The offline phase uses Hyperband via Keras Tuner on a large unlabeled 5-month dataset to derive tuned Bi-LSTM parameters, which are then employed during online anomaly detection across sliding windows. Evaluations on two real datasets (2d1a and 10d2a) employing Precision, Recall, and F1 show that window sizes in the 24–48 range yield the best balance, while the majority-detection criterion is generally ineffective. The study demonstrates that automated hyperparameter tuning can improve online contextual anomaly detection performance and offers a path toward more robust, domain-adaptive anomaly detectors for time-series data in smart environments.

Abstract

The exponential growth in the usage of Internet of Things in daily life has caused immense increase in the generation of time series data. Smart homes is one such domain where bulk of data is being generated and anomaly detection is one of the many challenges addressed by researchers in recent years. Contextual anomaly is a kind of anomaly that may show deviation from the normal pattern like point or sequence anomalies, but it also requires prior knowledge about the data domain and the actions that caused the deviation. Recent studies based on Recurrent Neural Networks (RNN) have demonstrated strong performance in anomaly detection. This study explores the impact of automatically tuned hyperparamteres on Unsupervised Online Contextual Anomaly Detection (UoCAD) approach by proposing UoCAD with Optimised Hyperparamnters (UoCAD-OH). UoCAD-OH conducts hyperparameter optimisation on Bi-LSTM model in an offline phase and uses the fine-tuned hyperparameters to detect anomalies during the online phase. The experiments involve evaluating the proposed framework on two smart home air quality datasets containing contextual anomalies. The evaluation metrics used are Precision, Recall, and F1 score.

Exploring the impact of Optimised Hyperparameters on Bi-LSTM-based Contextual Anomaly Detector

TL;DR

This work tackles contextual anomaly detection in IoT smart-home time-series by extending UoCAD with offline Bi-LSTM hyperparameter optimisation (UoCAD-OH). The offline phase uses Hyperband via Keras Tuner on a large unlabeled 5-month dataset to derive tuned Bi-LSTM parameters, which are then employed during online anomaly detection across sliding windows. Evaluations on two real datasets (2d1a and 10d2a) employing Precision, Recall, and F1 show that window sizes in the 24–48 range yield the best balance, while the majority-detection criterion is generally ineffective. The study demonstrates that automated hyperparameter tuning can improve online contextual anomaly detection performance and offers a path toward more robust, domain-adaptive anomaly detectors for time-series data in smart environments.

Abstract

The exponential growth in the usage of Internet of Things in daily life has caused immense increase in the generation of time series data. Smart homes is one such domain where bulk of data is being generated and anomaly detection is one of the many challenges addressed by researchers in recent years. Contextual anomaly is a kind of anomaly that may show deviation from the normal pattern like point or sequence anomalies, but it also requires prior knowledge about the data domain and the actions that caused the deviation. Recent studies based on Recurrent Neural Networks (RNN) have demonstrated strong performance in anomaly detection. This study explores the impact of automatically tuned hyperparamteres on Unsupervised Online Contextual Anomaly Detection (UoCAD) approach by proposing UoCAD with Optimised Hyperparamnters (UoCAD-OH). UoCAD-OH conducts hyperparameter optimisation on Bi-LSTM model in an offline phase and uses the fine-tuned hyperparameters to detect anomalies during the online phase. The experiments involve evaluating the proposed framework on two smart home air quality datasets containing contextual anomalies. The evaluation metrics used are Precision, Recall, and F1 score.
Paper Structure (13 sections, 2 equations, 2 figures, 7 tables)

This paper contains 13 sections, 2 equations, 2 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. UoCAD and UoCAD-OH
  4. Experiments and Results
  5. Datasets
  6. 2-days dataset (2d1a)
  7. 10-days dataset (10d2a)
  8. 5-months dataset (5M)
  9. Hyperparameter Optimisation
  10. Evaluation Metrics
  11. Experimental Setup
  12. Results and Discussion
  13. Conclusions

Figures (2)

  • Figure 1: Anomaly detection results of UoCAD-OH for different window sizes from the 2d1a dataset.
  • Figure 2: Anomaly detection results of UoCAD-OH for different window sizes from the 10d2a dataset.