ADL™: A Topic Model for Discovery of Activities of Daily Living in a Smart Home

Yu Chen, Tom Diethe, Peter A. FlachIJCAI 2016 · OpenReview

Abstract

We present an unsupervised approach for discovery of Activities of Daily Living (ADL) in a smart home. Activity discovery is an important enabling technology, for example to tackle the healthcare requirements of elderly people in their homes. The technique applied most often is supervised learning, which relies on expensive labelled data and lacks the flexibility to discover unseen activities. Building on ideas from text mining, we present a powerful topic model and a segmentation algorithm that can learn from unlabelled sensor data. The model has been evaluated extensively on datasets collected from real smart homes. The results demonstrate that this approach can successfully discover the activities of residents, and can be effectively used in a range of applications such as detection of abnormal activities and monitoring of sleep quality, among many others.

Introduction

Smart home environments are increasingly equipped with binary sensors such as motion detectors and door contacts. To leverage this data for healthcare applications, particularly for elderly care, systems must interpret these raw signals as meaningful Activities of Daily Living (ADL). Traditional supervised learning methods require extensive manual labeling, which is both costly and inflexible. We propose ADL™, an unsupervised topic model that learns activity patterns directly from unlabeled sensor data by drawing parallels between sensor event sequences and natural language processing.

Methodology: From Sensors to Topics

The ADL™ framework treats the stream of sensor data as a collection of "documents" composed of "words."

Data Segmentation and Vocabulary

A "word" is defined by a sensor ID and its state (e.g., a specific motion sensor turning ON). To create "documents," we employ a segmentation algorithm based on location changes. If a resident remains in a specific area for a duration exceeding a threshold ($t_{th}$), the sequence of events is bundled into a document.

The ADL™ Model

Building upon Latent Dirichlet Allocation (LDA), the ADL™ model incorporates both unigrams (individual sensor events) and bigrams (transitions between sensors). Bigrams are crucial because the sequence of movement—such as moving from a refrigerator to a stove—is often more indicative of a specific activity than isolated events.

The probabilistic graphical model of the ADL™ framework, illustrating the generation of unigrams and bigrams within documents.

The model uses Gibbs sampling to estimate the "topic" (activity) distribution for each document, effectively clustering related sensor patterns without prior labels.

Experimental Evaluation

The model was validated using real-world datasets from the CASAS project (hh122, hh120, and milan). While training was entirely unsupervised, ground-truth annotations were used to evaluate performance via the Fowlkes-Mallows (FM) Index.

ADL™ consistently outperformed standard LDA and Bigram Topic Models. For instance, on the milan dataset, ADL™ achieved an FM score of 0.6190, significantly higher than LDA's 0.5634. Furthermore, the model demonstrated a lower "Fragment Ratio," indicating its ability to group sensor events into cohesive activities rather than fragmented segments.

Visualization of discovered topics (8, 3, and 11) mapped onto a smart home floor plan, showing distinct spatial activity patterns discovered without labels.

Discussion and Key Insights

A key insight of this work is that location serves as a powerful proxy for intent. By capturing the "flow" of movement through bigrams, the model distinguishes between activities occurring in the same room. Beyond discovery, the model enables:

  • Data Compression: Reducing over 100,000 raw events into a few thousand activity documents.
  • Anomaly Detection: Identifying abnormal durations (e.g., a kitchen sensor active for hours) using z-score analysis.
  • Sub-topic Discovery: Using temporal clustering (K-means on start times) to refine broad topics like "Kitchen" into "Breakfast" or "Dinner."

Conclusion

ADL™ provides a robust, unsupervised solution for activity discovery in smart homes. While current segmentation relies on location transitions, future work will explore online learning and the direct integration of temporal features into the topic model to further enhance its predictive power and flexibility.

Summary updated

Abstract

We present an unsupervised approach for discovery of Activities of Daily Living (ADL) in a smart home. Activity discovery is an important enabling technology, for example to tackle the healthcare requirements of elderly people in their homes. The technique applied most often is supervised learning, which relies on expensive labelled data and lacks the flexibility to discover unseen activities. Building on ideas from text mining, we present a powerful topic model and a segmentation algorithm that can learn from unlabelled sensor data. The model has been evaluated extensively on datasets collected from real smart homes. The results demonstrate that this approach can successfully discover the activities of residents, and can be effectively used in a range of applications such as detection of abnormal activities and monitoring of sleep quality, among many others.
BibTeX
@inproceedings{chen2016adltm,
  author = {Yu Chen and Tom Diethe and Peter A. Flach},
  title = {{ADL™: A Topic Model for Discovery of Activities of Daily Living in a Smart Home}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2016},
  url = {https://openreview.net/forum?id=ByE0DmzuZS},
}

Loading related content...

Figure