Full Pipeline - Daily Physical Activity Monitoring: Adaptive Learning from Multi-Source Motion Sensor Data - Team 2

[thomasfire]

Team:

• Walter Tang (waltert3@illinois.edu)
• Egor Savkin (esavkin2@illinois.edu)
• Yihui Yu (yihuiy2@illinois.edu)
  Type: Full Pipeline
  Dataset: https://archive.ics.uci.edu/static/public/256/daily+and+sports+activities.zip
  Original paper:
• https://arxiv.org/abs/2405.16395
• https://proceedings.mlr.press/v248/zhang24a.html

Overview

This PR adds support for the Daily and Sports Activities (DSA) time-series dataset, an adaptive transfer model for multi-source time-series classification, a multi-class activity-recognition task, and a reproduction notebook with ablations and extensions based on the original paper.

Dataset (dsa.py)

1. Adds the DSADataset class in pyhealth.datasets.
2. Supports the Daily and Sports Activities (DSA) time-series dataset.
3. Prepares the dataset for downstream activity-recognition experiments in PyHealth.

Model (adaptive_transfer.py)

1. Adds the AdaptiveTransferModel class in pyhealth.models.
2. Implements an adaptive transfer approach for multi-source time-series classification.
3. Enables reproduction and extension of the method proposed in the original paper.

Task (dsa_activity_classification.py)

1. Adds the DSAActivityClassification task in pyhealth.tasks.
2. Supports multi-class activity recognition on the DSA dataset.

Example (dsa_adaptive_transfer_replication.ipynb)

1. Adds a Jupyter notebook for reproducing the original paper on the DSA dataset.
2. Includes our ablations and extensions for the adaptive transfer model.

Main Files for Review

1. pyhealth/datasets/dsa.py with tests tests/core/test_dsa.py
2. pyhealth/models/adaptive_transfer.py with tests tests/core/test_adaptive_transfer.py
3. pyhealth/tasks/dsa_activity_classification.py with tests tests/core/test_dsa_task.py
4. examples/dsa_adaptive_transfer_replication.ipynb