Abstract
Activity recognition aims to identify and predict human activities based on a series of sensor readings. In recent years, machine learning methods have become popular in solving activity recognition problems. A special difficulty for adopting machine learning methods is the workload to annotate a large number of sensor readings as training data. Labeling sensor readings for their corresponding activities is a time-consuming task. In practice, we often have a set of labeled training instances ready for an activity recognition task. If we can transfer such knowledge to a new activity recognition scenario that is different from, but related to, the source domain, it will ease our effort to perform manual labeling of training data for the new scenario. In this paper, we propose a transfer learning framework based on automatically learning a correspondence between different sets of sensors to solve this transfer-learning in activity recognition problem. We validate our framework on two different datasets and compare it against previous approaches of activity recognition, and demonstrate its effectiveness.
Introduction
Supervised machine learning has revolutionized activity recognition (AR), enabling systems to infer behaviors like "cooking" or "sleeping" from sensor data. However, these models are notoriously brittle. Traditionally, a model trained in one environment requires the exact same sensor configuration and activity labels to function in another. In real-world scenarios, homes differ in sensor types (e.g., motion sensors vs. RFID tags) and labeling conventions (e.g., "meal preparation" vs. "cooking"). Manually labeling data for every new environment is prohibitively expensive. This research addresses heterogeneous transfer learning, where the feature space, data distribution, and label space all vary between the source and target domains.
Methodology
The proposed framework bridges the source domain (labeled) and target domain (unlabeled) through a two-step mapping process:
- Sensor Mapping (Feature Space Transfer): The system identifies correspondences between different sensors by analyzing the "shape" of their data. It utilizes Jeffrey’s J-divergence to compare probability distributions of sensor values and Dynamic Time Warping (DTW) to align sequences with varying speeds or sampling rates. This allows the system to generate "pseudo-labels" for the target environment based on sensors that behave similarly to those in the source environment.
- Semantic Mapping (Label Space Transfer): To reconcile differing activity names, the authors leverage the Web as a source of common-sense knowledge. They use Normalized Google Distance (NGD) to calculate the semantic distance between terms. For instance, if "cooking" and "meal prep" frequently co-occur in web search results, the system infers they represent similar behaviors.
Final activity labels are assigned using a dynamic programming approach that minimizes expected loss, ensuring temporal consistency in the predicted activity sequences.
Experimental Results
The framework was evaluated on three real-world datasets: UvA (University of Amsterdam), MIT PLIA1 (MIT PlaceLab), and Intel (RFID-based morning activities).
Performance on UvA and Intel Datasets
The proposed method consistently outperformed unsupervised baselines. As shown in Figure 1, accuracy improved as the number of candidate labels ($K$) increased, reaching approximately 68.2% for UvA and 57.2% for Intel at $K=20$, significantly higher than the ~47% and ~42% achieved by unsupervised methods.
Cross-Dataset Transfer
A key strength of the framework is its ability to transfer knowledge between entirely different setups. When transferring from the large-scale MIT dataset to the Intel dataset, the system achieved a peak accuracy of 61.2% at $K=10$ (Figure 2). This is notable because the MIT dataset uses different sensor types and activity categories compared to the target domains.
Discussion and Conclusion
The research demonstrates that the statistical signature of a sensor is often more informative than its explicit identity. By treating the Web as a proxy for human common sense, the framework successfully relaxes assumptions regarding feature and label consistency.
However, limitations remain. The "relatedness" between source and target domains is difficult to quantify; if they are too disparate, the mapping may fail. Additionally, the current reliance on live web queries for NGD can introduce latency. Future work will explore more sophisticated probability summations and the potential use of pre-computed semantic embeddings to improve both accuracy and efficiency.
Summary updated