Contextual Anomaly Detection in Smart Homes Using Temporal Graph Based Distances

Amirhosein Bodaghi, Chris Nugent

Abstract: Identifying contextual anomalies in smart home environments is significant as it enables detailed health and medical applications, especially when the symptoms occur gradually instead of abruptly. Current methods often lack contextual awareness, limiting their effectiveness in accurately identifying anomalies within the complex activity patterns typical of smart home data. Our research addresses this gap by developing a method using graph edit distances (GED) to represent daily activity patterns. We first generated a synthetic dataset using a Markov model to mimic real-world smart home activity data. This process involved calculating transition probabilities and fitting various statistical distributions to the activity durations. Building upon this dataset, our anomaly detection approach involves constructing time series from GEDs data with multiple lag days and extracting residuals from them to avoid seasonality patterns. To evaluate our approach, we added synthetic anomalies to residuals by adjusting their values based on percentage changes across various sample sizes. We used two detection methods: Isolation Forest (IF) and Mahalanobis Distance (MD). IF was the most sensitive, with recall improving as the sample size increased, reaching 93.6% accuracy for 600 sample size at the highest percentage change. Conversely, MD’s effectiveness decreased with increasing sizes. In conclusion, this study presents an approach for detecting contextual anomalies in smart homes using temporal graphs, with IF achieving promising results for large anomaly sample sizes.

External IDs:doi:10.1007/978-3-031-77571-0_13