Label Space Driven Heterogeneous Transfer Learning With Web Induced Alignment

Sanatan SukhijaAAAI 2018 · DOI

Introduction

In the realm of smart home activity recognition, machine learning models often face significant hurdles due to the high cost of data labeling. Transfer learning offers a solution by leveraging knowledge from a source domain to a target domain. However, real-world applications are frequently heterogeneous, characterized by feature mismatches (different sensor types), label mismatches (varying activity naming conventions), and a lack of instance correspondences.

This paper introduces LSDHTL-WIA, a novel Heterogeneous Transfer Learning (HTL) algorithm designed to bridge these gaps. By utilizing "web-induced alignment"—specifically Normalized Google Distance (NGD) to identify semantic relationships between labels—the model successfully transfers knowledge across disparate smart home environments without the need for expensive grid searches typical of Latent Space Transformation (LST) methods.

Methodology

LSDHTL-WIA is categorized as a Feature Space Remapping (FSR) approach, which learns a direct transformation between source and target features. The framework is built upon three core pillars:

  • Web-Induced Alignment: The authors employ Normalized Google Distance (NGD) to calculate semantic similarity between source and target labels based on web-scale co-occurrence statistics. This creates a similarity matrix ($W$) that serves as a bridge between different label spaces.
  • Joint Optimization Objective: The algorithm minimizes a loss function comprising:
  • Reconstruction Loss: Preserves the original structural topology of both domains.
  • Semantic Co-alignment: Uses the NGD-based matrix to penalize distances between source and target instances with semantically related labels, effectively "pulling" related activities together in the projected space.
  • Regularization: Prevents overfitting to the source training data.
  • Alternating Minimization: An iterative approach is used to solve the non-convex objective function, ensuring convergence by fixing two variables while optimizing the third.

Experimental Setup

The model was rigorously evaluated using the CASAS horizon house datasets, which include data from three single-resident smart homes: hh102, hh113, and hh118. The researchers conducted six cross-domain activity recognition tasks (e.g., training on hh102 and testing on hh113) to test the robustness of the transfer learning process.

Results and Discussion

LSDHTL-WIA was compared against standard baselines (Random Forests, SVMs) and state-of-the-art transfer methods, including HFA, HeMap, and SHDA-RF.

Performance comparison of LSDHTL-WIA against baseline and state-of-the-art transfer learning methods on CASAS horizon house datasets.

As shown in the results table, LSDHTL-WIA consistently outperformed all competing methods across all six transfer tasks. Notably, in the hh102 → hh113 task, the proposed method achieved a mean error of 25.19%, a significant improvement over the SVM baseline (34.58%) and the previous best-performing transfer method, SHDA-RF (26.37%). The low standard deviations across tasks further underscore the method's robustness.

Conclusion

The success of LSDHTL-WIA highlights the power of the "wisdom of the web" in bypassing rigid ontologies for semantic alignment. By treating the problem as a Feature Space Remapping task, the authors avoid the computational overhead of finding latent subspaces while maintaining structural integrity through reconstruction loss.

Despite its strengths, the method currently requires some labeled data in the target domain to establish semantic alignment, leaving the "cold start" problem (zero labeled target data) as a challenge for future research. Expanding the framework to incorporate more diverse external knowledge sources could further refine the alignment process.

Summary updated

BibTeX
@inproceedings{sukhija2018label,
  author = {Sanatan Sukhija},
  title = {{Label Space Driven Heterogeneous Transfer Learning With Web Induced Alignment}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.1609/aaai.v32i1.12166},
}
Figure