Research map / Label-efficient and robust learning / Recognition and long-tail
Domain-Informed Label Fusion Surpasses LLMs in Free-Living Activity Classification (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FuSE-MET, which merges noisy free-living activity labels by clustering a lambda-weighted fusion of BERT word embeddings and MET intensity values, then trains standard classifiers on the reduced label space.
Problem
Human activity recognition in free-living settings suffers from noisy, inconsistently worded, sparse user-provided activity labels and an undefined activity vocabulary, making classification hard.
Approach
Label merging is framed as clustering of label feature vectors that combine BERT word embeddings (semantic meaning) with MET values (physical intensity) via a lambda-weighted fusion, followed by k-means, to reduce label complexity before classification.
Key results
- With K=3 clusters, FuSE-MET with the best classifier reaches F1 0.77 (accuracy 0.68), versus 0.59 for the GPT4+CNN baseline in the reported table.
- Optimal lambda (domain knowledge weight) lies in the range 0.2-0.5 for the reported settings.
- The no-fusion 1NN baseline reaches F1 0.18 with 36 distinct labels.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add MET-value fusion (lambda-weighted BERT + MET label embedding) before clustering vs. GPT-4-based label fusion without MET | label merging for activity classifiers · Free-living HAR (cardiovascular study) | F1 | Helped |
| Add label fusion (FuSE-MET) vs. no-fusion with 36 distinct labels | 1NN classifier · Free-living HAR (cardiovascular study) | F1 | Helped |
| Reduce label space to K=3 clusters vs K=2 clusters | FuSE-MET with XGB · Free-living HAR (cardiovascular study) | F1: 0.77 vs 0.65 | Mixed |
| Set domain coefficient lambda in 0.2-0.5 range (MET weighting) (ablation) | FuSE-MET · Free-living HAR (cardiovascular study) | F1 | Helped |
| Use simpler classifiers (1NN, SVM) on FuSE-MET merged labels | FuSE-MET label space · Free-living HAR (cardiovascular study) | F1 | No clear effect |
Surprising findings
- Simpler classifiers such as 1NN and SVM perform better than complex models once labels are fused, which the authors attribute to reduced label complexity.
Limitations
- Evaluated on a single clinical study dataset of cardiovascular patients, with a small reported set of results.
- Table results are mixed: K=2 accuracy for SVM is high while F1 is low, and some GPT-4 fused baselines are close to FuSE-MET.
Experimental setup
Datasets: clinical study data from patients with cardiovascular disease (accelerometer and gyroscope smartphone data with user-ann
Benchmarks: in-house free-living cardiovascular patient activity dataset
Metrics: accuracy precision recall F1
Baselines: GPT-4-based label fusion no-fusion model SVM XGB 1NN DANets CNN
Builds on: BERT
Closest papers by meaning
- Limitations in Employing Natural Language Supervision for Sensor-Based Human Activity Recognition - And Ways to Overcome Them (AAAI 2025 · similarity 0.84)
- Generalizable Sensor-Based Activity Recognition via Categorical Concept Invariant Learning (AAAI 2025 · similarity 0.83)
- HARDVS: Revisiting Human Activity Recognition with Dynamic Vision Sensors (AAAI 2024 · similarity 0.82)
- SeFAR: Semi-supervised Fine-grained Action Recognition with Temporal Perturbation and Learning Stabilization (AAAI 2025 · similarity 0.81)
- Self-Paced Unified Representation Learning for Hierarchical Multi-Label Classification (AAAI 2024 · similarity 0.81)
- MoPFormer: Motion-Primitive Transformer for Wearable-Sensor Activity Recognition (NeurIPS 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.