IoT2Vec: Identification of Similar IoT Devices via Activity Footprints

We consider a smart home or smart office environment with a number of IoT devices connected and passing data between one another. The footprints of the data transferred can provide valuable information about the devices, which can be used to (a) identify the IoT devices and (b) in case of failure, to identify the correct replacements for these devices. In this paper, we generate the embeddings for IoT devices in a smart home using Word2Vec, and explore the possibility of having a similar concept for IoT devices, aka IoT2Vec. These embeddings can be used in a number of ways, such as to find similar devices in an IoT device store, or as a signature of each type of IoT device. We show results of a feasibility study on the CASAS dataset of IoT device activity logs, using our method to identify the patterns in embeddings of various types of IoT devices in a household.

💡 Research Summary

The paper introduces IoT2Vec, a framework that leverages activity logs from smart homes or offices to generate vector embeddings for IoT devices using the Word2Vec algorithm. The authors treat each device identifier as a “word” and a sequence of device state changes occurring within a predefined time window as a “sentence.” By extracting only meaningful state transitions (e.g., ON↔OFF) and discarding noisy sensors (such as light or gyro sensors), they construct a corpus of sentences that reflect the temporal co‑activation patterns of devices.

Four lemmas underpin the approach: (1) devices of the same type exhibit recurring usage patterns; (2) these patterns are distinctive enough to differentiate device categories; (3) the time of activation and physical location add discriminative information; and (4) such patterns can be captured by word‑embedding models like Word2Vec or GloVe. Based on these premises, the authors outline an eight‑step pipeline: data cleaning, transition extraction, session definition (using a configurable “session gap”), sentence construction, Word2Vec training (typically 100‑dimensional vectors), similarity computation, dimensionality reduction with PCA‑t‑SNE, and visual analysis of clusters.

An algorithm is presented for classifying an unknown device: generate its embedding from its log, compute cosine similarity with stored embeddings of known device types, and assign the type if the highest similarity exceeds a threshold. This simple yet effective procedure enables automatic identification of replacement devices, similarity‑based search in IoT marketplaces, and location‑based classification (e.g., recognizing a “kitchen” or “toilet” area from the devices that tend to fire together).

The experimental evaluation uses the CASAS Kyoto dataset, which contains two years of sensor data from a household with multiple residents. The authors select a subset of sensors (motion, door, item, shake, fan, experimental switch) and create sessions with three different gaps: 10 seconds, 60 seconds, and 600 seconds. Word2Vec embeddings are trained on these sessions, and t‑SNE visualizations reveal coherent clusters. For a 60‑second gap, the freezer door sensor (D008) clusters closely with kitchen motion sensors (M015‑M017) and the fridge door sensor (D009), reflecting the real‑world co‑occurrence of opening the freezer or fridge while moving in the kitchen. Similarly, sensors located near the toilet (M038‑M041, D005‑D006) form a tight cluster across all gap settings, indicating that spatial proximity translates into embedding similarity regardless of temporal granularity.

The results demonstrate that (a) embeddings capture contextual relationships driven by both functional usage and physical placement, (b) the choice of session gap influences the granularity of observed relationships but does not destroy core spatial clusters, and (c) cosine similarity between embeddings can reliably infer device type for new, unseen devices.

Limitations acknowledged include dependence on the quality and completeness of logs, sensitivity to the session‑gap parameter, and the fact that experiments are confined to a single household dataset, limiting generalizability. The authors propose future work such as incorporating data from multiple homes, developing adaptive session segmentation methods, integrating sequential models (LSTM, Transformers) to capture longer‑range dependencies, and building end‑to‑end recommendation systems for device replacement or procurement.

In conclusion, IoT2Vec shows that purely behavioral data—without any metadata about device specifications—can be transformed into meaningful low‑dimensional representations. These embeddings enable practical applications like automated device identification, similarity‑based search, and context‑aware services, opening a new research direction at the intersection of natural language processing and Internet‑of‑Things analytics.