LabelForest: Non-Parametric Semi-Supervised Learning for Activity Recognition

Yuchao Ma, Hassan GhasemzadehAAAI 2019 · DOI

Introduction

Personalizing activity recognition (AR) models for wearable devices, such as smartwatches and fitness trackers, presents a significant challenge: human movement patterns are highly individual. Achieving high accuracy typically requires extensive labeled data from each specific user. However, expecting users to manually label hundreds of activity instances is impractical, leading to a "cold start" problem in personalized AR.

Semi-supervised learning (SSL) offers a potential solution by leveraging a small "seed" of labeled data to automatically annotate a larger unlabeled dataset. Despite this, existing SSL methods often struggle with wearable sensor data due to hyper-parameter sensitivity (e.g., tuning $k$ in $k$-NN), rigid statistical distribution assumptions, and extreme sensitivity to skewed seed labels (where one activity has significantly more initial examples than others).

System Architecture

To address these limitations, the authors propose LabelForest, a non-parametric framework designed to accurately label massive amounts of sensor data starting from as little as one labeled example per activity. The system follows a two-stage pipeline that avoids pre-defined thresholds and distribution assumptions.

The LabelForest system architecture, illustrating the flow from raw input and feature extraction to the Greedy Spanning Forest and final model training.

As shown in the architecture above, the process begins with feature extraction from wearable sensors. These features, along with initial seed labels, are fed into the Greedy Spanning Forest (GSF) algorithm. After the forest is grown, a sample filtering stage ensures only high-confidence labels are used to train the final supervised model (e.g., an SVM) for real-time recognition.

The LabelForest Algorithm

Greedy Spanning Forest (GSF)

The core of LabelForest is the GSF algorithm, which treats label propagation as a territory expansion process. Starting with seed labels as "roots," the algorithm grows trees by iteratively adding the most similar unlabeled data points.

Step-by-step evolution of the Greedy Spanning Forest: (a) initial growth, (b-c) detection of conflicting edges and margin calculation, and (d) the final partitioned forest.

A critical innovation in GSF is the Conflicting Edge logic. When two trees representing different activities (e.g., "Walking" vs. "Running") approach each other in the feature space, the algorithm identifies a "conflicting edge" (the red line in Figure 3a). It calculates a margin based on the similarity at this boundary and stops a tree's growth if the next potential point is less similar to the root than the points at the boundary. This data-driven approach prevents "label leakage" across decision boundaries without requiring manual hyper-parameter tuning.

Silhouette-based Filtering

Following the GSF phase, LabelForest applies a silhouette-based filtering mechanism. Not all assigned labels are equally reliable; therefore, the system calculates a confidence score for each point based on its proximity to its assigned cluster versus the next nearest cluster. Low-confidence labels are pruned, resulting in a high-quality, augmented training set that prioritizes precision over sheer quantity.

Experimental Evaluation

The researchers evaluated LabelForest across three diverse datasets: HART (daily activities), SmartSock (lower-limb activities), and Phone (smartphone-based accelerometer data).

Key findings include:

  • High Labeling Accuracy: LabelForest achieved a labeling accuracy of 90.1%, even when starting with highly skewed seed distributions.
  • Significant Performance Gains: Compared to standard self-training methods, LabelForest demonstrated a 56.9% improvement on balanced data and a 175.3% improvement on unbalanced data.
  • Robustness to Minimal Seeds: With only one seed sample per activity, the final model maintained over 71.9% accuracy across all datasets. Even when seeds were skewed (e.g., one class having 8 samples and others only 1), accuracy remained stable at over 73.1%.

Conclusion and Future Work

LabelForest effectively mitigates the cold-start problem in personalized activity recognition by using a non-parametric, graph-based approach that is robust to skewed data and requires no manual thresholding. While currently an offline batch process, future work aims to adapt LabelForest for incremental learning, allowing wearable systems to refine labels in real-time as users generate new data. Additionally, integrating this graph-based logic with deep learning architectures presents a promising avenue for further research.

Summary updated

BibTeX
@inproceedings{ma2019labelforest,
  author = {Yuchao Ma and Hassan Ghasemzadeh},
  title = {{LabelForest: Non-Parametric Semi-Supervised Learning for Activity Recognition}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.33014520},
}
Figure